Use the binomial distribution when you count successes in a fixed number n of independent trials, each with the same probability p. Use the Poisson distribution when you count events in a stretch of time, length or area with a known average rate and no fixed upper limit. Use the normal distribution for a continuous measurement such as a diameter or a voltage, or as an approximation to a binomial count when n is large.
That is the whole decision in three sentences. The hard part in an exam is reading a paragraph of engineering context and seeing which of the three it describes, then setting up the probability so that "at least", "at most" and "more than" come out right. This guide works through six examples in rising difficulty, all with our own numbers.
The three models side by side
| Binomial B(n, p) | Poisson Po(λ) | Normal N(μ, σ²) | |
|---|---|---|---|
| What is counted or measured | successes in n trials | events in an interval | a continuous quantity |
| Possible values | 0, 1, …, n | 0, 1, 2, … (no upper limit) | any real number |
| Formula | P(X = k) = C(n, k) pᵏ (1 − p)ⁿ⁻ᵏ | P(X = k) = e^(−λ) λᵏ / k! | use z = (x − μ)/σ and tables |
| Mean | np | λ | μ |
| Variance | np(1 − p) | λ | σ² |
| Typical wording | "a sample of 15", "each with probability 0.08" | "on average 4.2 per hour", "0.6 defects per metre" | "diameters are normally distributed with mean…" |
Two facts in this table do a lot of work. A Poisson variable has mean equal to variance, so if the question gives you a count whose mean and variance are roughly equal, Poisson is a strong candidate. A binomial variable can never exceed n, so if there is no natural ceiling on the count, it is not binomial.
If you are taking a probability or statistics module and want a structured route through it, our online probability and statistics tutoring page explains how sessions with Ajay Vatsyayan work.
How to recognise which one you need
Ask these questions in order.
- Is the quantity measured rather than counted? Lengths, masses, times, voltages: normal (or another continuous model the question names).
- Is there a fixed number of trials, each a yes/no outcome with the same probability, independent of each other? Binomial.
- Is it a count of events over time, length, area or volume, given as an average rate? Poisson. Rescale the rate to the interval asked about.
- Is it binomial with n large and p small (roughly n ≥ 50 and np ≤ 5)? You may approximate by Poisson with λ = np.
- Is it binomial with n large and p not close to 0 or 1 (np and n(1 − p) both above about 5)? You may approximate by normal with μ = np and σ² = np(1 − p), using a continuity correction.
Worked example 1: a binomial sample
A supplier's resistors are defective with probability 0.08, independently. An inspector tests a sample of 15. Find the probability that at most one is defective, and the probability that at least two are.
Let X be the number defective. Fixed n = 15, yes/no outcome, constant p = 0.08: X ~ B(15, 0.08).
P(X = 0) = 0.92¹⁵ = 0.2863
P(X = 1) = 15 × 0.08 × 0.92¹⁴ = 0.3734
P(X ≤ 1) = 0.2863 + 0.3734 = 0.6597
P(X ≥ 2) = 1 − P(X ≤ 1) = 0.3403
Mean np = 1.2 and variance np(1 − p) = 1.104, so the standard deviation is about 1.05. Notice the complement step: "at least two" is everything except 0 and 1. Writing 1 − P(X ≤ 2) here is the most common single slip in this topic.
Worked example 2: a Poisson rate and rescaling
A telecom exchange records dropped calls at an average rate of 4.2 per hour, at random and independently. Find (a) the probability of no dropped calls in an hour, (b) the probability of at least three in an hour, (c) the probability of at most one in a 30-minute period.
X ~ Po(4.2) for one hour.
(a) P(X = 0) = e^(−4.2) = 0.0150
(b) P(X ≤ 2) = e^(−4.2)(1 + 4.2 + 4.2²/2) = 0.01500 + 0.06298 + 0.13226 = 0.21024, so P(X ≥ 3) = 0.7898.
(c) Half an hour has rate λ = 4.2 × ½ = 2.1. With Y ~ Po(2.1):
P(Y ≤ 1) = e^(−2.1)(1 + 2.1) = 0.3796
The rate belongs to the interval. Students who keep λ = 4.2 for part (c) get the wrong answer with a correct-looking method, and lose almost all the marks.
Worked example 3: Poisson as an approximation to binomial
A pipeline has 1,200 welds. Each weld is faulty with probability 0.0025, independently. Find the probability of at least five faulty welds.
The exact model is B(1200, 0.0025). Computing C(1200, k) by hand is unpleasant, and n is large while p is small, so approximate by Poisson with λ = np = 3.
P(X ≤ 4) ≈ e^(−3)(1 + 3 + 9/2 + 27/6 + 81/24) = 0.04979 + 0.14936 + 0.22404 + 0.22404 + 0.16803 = 0.81526
P(X ≥ 5) ≈ 0.1847
The exact binomial answer is 0.1845, so the approximation is accurate to three decimal places. In a written answer, state the conditions (n large, p small) and the value λ = np before you use it. Examiners usually award a mark for that justification.
Worked example 4: a normal tolerance band
Shaft diameters are normally distributed with mean 25.00 mm and standard deviation 0.04 mm. The tolerance band is 24.93 mm to 25.06 mm. (a) What proportion of shafts is within tolerance? (b) What symmetric band 25.00 ± c mm contains 99% of shafts?
(a) Standardise each end:
z₁ = (24.93 − 25.00)/0.04 = −1.75
z₂ = (25.06 − 25.00)/0.04 = 1.50
P(−1.75 < Z < 1.50) = Φ(1.50) − Φ(−1.75) = 0.9332 − 0.0401 = 0.8931
So about 89.3% of shafts pass.
(b) We need P(−c/0.04 < Z < c/0.04) = 0.99, so c/0.04 is the z-value with 0.995 to its left, which is 2.5758.
c = 2.5758 × 0.04 = 0.103 mm
The band is roughly 24.897 mm to 25.103 mm. A frequent error is dividing by the variance 0.0016 instead of the standard deviation 0.04.
Worked example 5: normal approximation with a continuity correction
Of components put through a burn-in test, 10% fail, independently. A batch of 400 is tested. Estimate the probability that 30 or fewer fail.
X ~ B(400, 0.1). Here np = 40 and n(1 − p) = 360, both large, so use a normal approximation with μ = 40 and σ² = 400 × 0.1 × 0.9 = 36, σ = 6.
"30 or fewer" means the bars at 0, 1, …, 30. The bar at 30 runs from 29.5 to 30.5, so the continuous version is P(Y ≤ 30.5).
z = (30.5 − 40)/6 = −1.583
P(X ≤ 30) ≈ Φ(−1.583) = 0.0567
The exact binomial value is 0.0524. Without the continuity correction you would get Φ(−1.667) = 0.0478, which is further off. The correction shifts by ½ in the direction that includes the boundary value: ≤ 30 becomes < 30.5, ≥ 30 becomes > 29.5, < 30 becomes < 29.5.
Worked example 6: a Poisson inside a binomial
Defects in a coated cable occur at random at an average of 0.6 per metre. The cable is cut into 2 m lengths, and a length is accepted if it has at most one defect. Ten lengths are cut. Find the probability that at least eight are accepted.
This question uses two models, one inside the other.
Step 1, one length. Defects in 2 m: D ~ Po(1.2).
q = P(D ≤ 1) = e^(−1.2)(1 + 1.2) = 0.6626
Step 2, ten lengths. Each length is accepted or not, independently, with the same probability q, so the number accepted A ~ B(10, 0.6626).
P(A ≥ 8) = C(10, 8) q⁸ (1 − q)² + C(10, 9) q⁹ (1 − q) + q¹⁰ = 0.2898
Questions like this appear in quality-control and reliability sections of engineering papers because they test whether you can name the random variable at each stage. Write "let D be…" and "let A be…" in words. Keep q to four or more significant figures, because it is raised to the tenth power.
Where students lose marks
- Wrong complement. P(X ≥ 2) = 1 − P(X ≤ 1), not 1 − P(X ≤ 2). Write the set of values you are excluding.
- Not rescaling λ. A rate per hour is not a rate per 30 minutes. Multiply the rate by the length of the new interval.
- Using the variance in z. z = (x − μ)/σ. If the question says "variance 0.0016", take the square root first.
- Missing or reversed continuity correction. Only use it when a discrete count is approximated by a continuous normal, and draw the bar to decide whether to add or subtract ½.
- Poisson approximation without justification. State "n is large and p is small, so X is approximately Po(np)". Using it when p = 0.3 is a modelling error.
- Assuming independence the context breaks. Sampling without replacement from a small batch is not binomial. If 3 of 20 items are faulty and you take 5, the hypergeometric model applies.
- Rounding too early. Carry four significant figures through intermediate steps, as in example 6.
Practice set
- In a data stream, 10% of sensor readings are outliers, independently. In 20 readings, find the probability of exactly three outliers.
- Typing errors occur at an average of 1.5 per page. Find the probability of no errors in a two-page report.
- Battery life is normal with mean 480 hours and standard deviation 30 hours. Find the probability that a battery lasts more than 525 hours.
- A batch has 2,000 chips, each defective with probability 0.001. Using a Poisson approximation, find the probability of at most one defective chip.
- A process succeeds with probability 0.4 per trial. In 250 trials, use a normal approximation to estimate the probability of at least 110 successes.
Answers. (1) C(20, 3)(0.1)³(0.9)¹⁷ = 0.1901. (2) λ = 3, e^(−3) = 0.0498. (3) z = 1.5, probability 0.0668. (4) λ = 2, 3e^(−2) = 0.4060 (exact binomial 0.4059). (5) μ = 100, σ = √60 ≈ 7.746, P(Y ≥ 109.5) with z ≈ 1.226, about 0.110.
What this leads to
These three distributions are the base for most of a first statistics course. The normal distribution and the way a binomial count becomes approximately normal lead to the central limit theorem, confidence intervals and hypothesis tests. The Poisson distribution leads to the exponential distribution for waiting times, Poisson processes and queueing models, which appear in networking, operations research and reliability engineering. The fact that Poisson probabilities add to 1 is the exponential series e^λ = 1 + λ + λ²/2! + …, which is one reason the convergence tests for series matter outside pure maths.
If you met these distributions at school, our guides to the IB Poisson distribution and IB probability distributions are a quick refresher before the university version. Students resitting a paper may also find our plan for clearing an engineering maths back paper useful, and students moving from JEE preparation can read what changes in IIT first-year maths.
How Maths Bodhi helps
We work through your own course's problem sheets and past papers, check your set-up of each random variable, and give you extra questions of the kinds you find hard. Engineering students can also see our engineering mathematics sessions, and other courses are listed on our university maths hub. We explain, practise and review your own attempts. We don't complete graded assignments, quizzes or exams.
Home tuition is available in Gurgaon (Gurugram). Students elsewhere in India and abroad learn with us online. You can book a free demo class.
Written and checked by the Maths Bodhi team (November 2026).
Frequently asked questions
- What is the difference between the Poisson and binomial distributions?
- Binomial counts successes in a fixed number of trials with probability p each; Poisson counts events in an interval at an average rate, with no upper limit.
- When can I use the Poisson approximation to the binomial?
- When n is large and p is small, roughly n ≥ 50 and np ≤ 5, using λ = np.
- When do I use a continuity correction?
- Whenever you approximate a discrete count, such as a binomial, by a continuous normal distribution.
- Why are the mean and variance of a Poisson distribution equal?
- Both work out to λ from the formula; it also follows from the binomial np(1 − p) as p → 0 with np fixed at λ.
- Is the number of defective items in a sample binomial or Poisson?
- Binomial if the sample size is fixed and each item is defective independently with the same probability; Poisson approximates it when n is large and p is small.
- What is a real-life example of a Poisson distribution in engineering?
- The number of faults per kilometre of cable, network packets arriving per second, or component failures per month.
