Lesson 8 of 8 · 20 min

Modelling with exponential random variables

Find quantiles and medians of exponential waiting times, find λ from given information, use the memoryless property, and evaluate whether an exponential model suits the context.

What you will be able to do

  • Find the quantile qq with P(T≤q)=pP(T \leq q) = p, including the median ln⁡(2)λ\frac{\ln(2)}{\lambda}.
  • Find λ\lambda from a mean, a median or a given probability.
  • Use the memoryless property P(T>s+t∣T>s)=P(T>t)P(T > s + t \mid T > s) = P(T > t).
  • Model waiting times and lifetimes, with and without technology, and judge the model's assumptions.

Probability questions ask "how likely is a wait of more than 10 minutes?". Quantile questions turn that round: "how long a wait covers 90 % of customers?" or "what warranty period will at most 5 % of products fail within?". For an exponential variable both directions are short calculations with e−λte^{-\lambda t} and logarithms.

Interactive

The median of an exponential distribution

Time between calls with λ=0.25\lambda = 0.25 per minute. Half the area lies to the left of the median, 4ln⁡(2)≈2.774\ln(2) \approx 2.77 minutes, which is well below the mean of 44 minutes. The long right tail pulls the mean up.

Key idea

Quantiles

The pp quantile qq satisfies P(T≤q)=pP(T \leq q) = p:

1−e−λq=p  ⟹  e−λq=1−p  ⟹  q=−ln⁡(1−p)λ1 - e^{-\lambda q} = p \;\Longrightarrow\; e^{-\lambda q} = 1 - p \;\Longrightarrow\; q = -\frac{\ln(1 - p)}{\lambda}

The median is the 0.50.5 quantile, ln⁡(2)λ≈0.693λ\frac{\ln(2)}{\lambda} \approx \frac{0.693}{\lambda}. It is always less than the mean 1λ\frac{1}{\lambda}, because P(T<mean)=1−e−1≈0.632P(T < \text{mean}) = 1 - e^{-1} \approx 0.632.

Solving an exponential equation

  1. Write the probability statement

    P(T≤q)=pP(T \leq q) = p or P(T>q)=pP(T > q) = p, whichever the question describes.

  2. Replace it with exponentials

    1−e−λq=p1 - e^{-\lambda q} = p, or e−λq=pe^{-\lambda q} = p.

  3. Isolate the exponential, then take logarithms

    Then divide by −λ-\lambda. To find λ\lambda instead, divide by −q-q.

  4. Answer in context

    Exact in Paper 1; rounded with units in Paper 2.

Worked example

Median, tail and 90th percentile of a lifetime

  • Complex familiar

The lifetime of a type of sensor is exponential with a median of 2000 hours.

a) Determine the exact value of λ\lambda.

b) Determine the probability that a sensor lasts more than 5000 hours, correct to three decimal places.

c) Determine the time by which 90 % of sensors have failed, to the nearest hour.

3 steps

Try it yourself first, then check the working one step at a time.

Method

  1. From a median mm: λ=ln⁡(2)m\lambda = \frac{\ln(2)}{m}.

  2. P(T>t)=e−λtP(T > t) = e^{-\lambda t}.

  3. For the pp quantile, solve 1−e−λq=p1 - e^{-\lambda q} = p.

Watch out for

  • Treating the median as the mean and using λ=12000\lambda = \frac{1}{2000}.

  • Solving e−λq=0.9e^{-\lambda q} = 0.9 for the 90th percentile: that gives the time by which 10 % have failed.

The time, TT years, until a smoke alarm battery fails is exponential with λ=0.04\lambda = 0.04. Determine the exact median lifetime.

Reads asYour answer shows here as maths.

No memory. Suppose the waiting time between calls is exponential and you have already waited ss minutes. The chance of waiting at least another tt minutes is

P(T>s+t∣T>s)=P(T>s+t)P(T>s)=e−λ(s+t)e−λs=e−λt=P(T>t)P(T > s + t \mid T > s) = \frac{P(T > s + t)}{P(T > s)} = \frac{e^{-\lambda(s + t)}}{e^{-\lambda s}} = e^{-\lambda t} = P(T > t)

The time already waited makes no difference. For a component, this means a used one is as good as a new one: it does not wear out. That is a strong assumption, and a good thing to question in an evaluation.

The lifetime of an LED driver is exponential with a mean of 1500 hours. A driver has already run for 1000 hours. Determine the exact probability that it runs for at least another 500 hours.

Reads asYour answer shows here as maths.

Worked example

Choosing a warranty period

  • Complex unfamiliar
  • Technology active

The lifetime of a cordless drill is modelled as exponential with a mean of 6 years. The manufacturer wants to offer a warranty, in whole months, under which at most 5 % of drills fail.

a) Determine the longest warranty the manufacturer can offer.

b) Evaluate the suitability of the exponential model for this decision.

5 steps

Try it yourself first, then check the working one step at a time.

Method

  1. Define TT with units and λ\lambda.

  2. Turn the requirement into an inequality in P(T≤w)P(T \leq w).

  3. Solve with logarithms, keeping track of the inequality's direction.

  4. Round in the direction the context requires.

  5. Evaluate the constant-rate assumption.

Watch out for

  • Rounding 3.69 months up to 4 months, which breaks the 5 % limit.

  • Solving e−w/6=0.05e^{-w/6} = 0.05, which finds when 95 % have failed.

Without technology

Leave answers as e−1.5e^{-1.5}, 1−e−21 - e^{-2}, 5ln⁡(2)5\ln(2) or −ln⁡(0.9)λ-\frac{\ln(0.9)}{\lambda}. Simplify logarithms: −ln⁡(12)=ln⁡(2)-\ln\left(\frac{1}{2}\right) = \ln(2) and −ln⁡(0.1)=ln⁡(10)-\ln(0.1) = \ln(10).

With technology

Most graphics calculators have no exponential distribution menu. Evaluate e−λte^{-\lambda t} directly, or integrate the pdf numerically, ∫abλe−λt dt\int_a^b \lambda e^{-\lambda t}\,dt, and use the solve function for quantiles. Write down what you evaluated.

Note

When is an exponential model reasonable?

The model assumes events occur independently, one at a time, at a constant average rate. Arrivals at a shop are not constant across the day; a car battery wears out, so an old one is more likely to fail than a new one. Say which assumption fails and what it would do to the answers: a model that ignores wear-out underestimates failures late in life.

Exam tip

Define the random variable

Begin with a sentence such as "Let TT be the time in minutes until the next call, T∼exponential(λ=0.25)T \sim \text{exponential}(\lambda = 0.25)." It fixes the units, earns a communication mark, and stops you mixing λ\lambda with the mean.

Summary

  • The pp quantile is q=−ln⁡(1−p)λq = -\frac{\ln(1 - p)}{\lambda}; the median is ln⁡(2)λ\frac{\ln(2)}{\lambda}, less than the mean.

  • Find λ\lambda from a mean (λ=1μ\lambda = \frac{1}{\mu}), a median (λ=ln⁡(2)m\lambda = \frac{\ln(2)}{m}) or a probability, using logarithms.

  • Memoryless: P(T>s+t∣T>s)=P(T>t)P(T > s + t \mid T > s) = P(T > t).

  • Define TT with units and its parameter before you calculate.

  • An exponential model needs a constant rate of independent events; question it when things wear out or rates change.

Finished the lesson?

Try the 2 quick checks above and the lesson completes itself, or mark it complete now.

Lesson 8 of 80% read

Save your progress with a free account