Probability Distributions: Normal, Binomial, and Poisson

The Three Kinds of Randomness You'll Actually Meet

On CBSE Class 10 result day, a school's marks list has a shape. Very few students score near 0, very few score near 100, and a thick cluster sits somewhere in the middle — a hump that tapers off symmetrically on both sides. Nobody designed the exam to produce that shape on purpose; it simply emerges whenever a score is the sum of many small, mostly independent effects (how well you knew each chapter, how each question happened to be phrased, how alert you were that morning). That hump has a name, and it appears everywhere from height distributions to stock returns: the Normal distribution.

Now switch contexts. A bowler runs in to bowl a fixed over of 6 balls. Each ball is either a six or it isn't — two outcomes, a fixed number of attempts, and (roughly) the same underlying skill on every ball. If you want to know the chance of exactly 2 sixes in that over, you are no longer asking about a smooth hump — you're counting a specific number of successes out of a fixed number of independent tries. That's a Binomial distribution problem.

Now stretch the same idea to an extreme. Instead of 6 balls, imagine 50,000 UPI transactions passing through a payment gateway in one minute, each with only a tiny, tiny chance of failing due to a network timeout. You can't realistically track 50,000 individual "trials" — but you can say that failures happen at some average rate, say 4 per minute, and ask: what's the chance exactly 2 failures happen in the next minute? This is a different shape of question again, and it belongs to the Poisson distribution.

All three of these are answers to the same underlying question — "how likely is each possible outcome of a random process?" — but the process looks different in each case, and using the wrong tool gives you a wrong (and sometimes absurd) answer. This chapter builds all three from first principles, shows precisely how they connect to each other, and gives you the judgment to pick the right one under exam pressure.

Random Variables: Discrete vs Continuous

A random variable is just a number that comes out of a random process — the number of sixes in an over, the number of dropped calls in an hour, a student's marks out of 100. Random variables split into two families that behave very differently:

• Discrete random variables take countable values — 0, 1, 2, 3, ... You can list them. Number of sixes, number of failed transactions, number of defective chips in a batch — all discrete. For these, P(X = k) is a genuine, meaningful probability, and the full list of these probabilities is called a probability mass function (PMF).
• Continuous random variables take any value in a range — 62.3 marks, 62.31 marks, 62.314 marks, and so on without end. Height, weight, exam marks treated as a continuous scale, reaction time — all continuous. For these, we use a probability density function (PDF), and probability is read off as area under the curve over an interval, never as the height of the curve at one exact point.

Binomial and Poisson are discrete. Normal is continuous. Keeping this straight prevents a mistake we'll return to later: asking for "the probability a student scores exactly 210.0000 marks" under a Normal model, which is a nonsensical question with a real conceptual reason behind why it's nonsensical.

The Binomial Distribution: Counting Successes in Fixed Trials

A single yes/no trial with a fixed success probability p is called a Bernoulli trial. "Success" happens with probability p, "failure" with probability q = 1 - p. A Binomial distribution is what you get when you repeat a Bernoulli trial n times, independently, and count the total number of successes. Four conditions must hold simultaneously:

1. A fixed number of trials, n, decided before you start.
2. Each trial has only two outcomes.
3. The success probability p is the same on every trial.
4. Trials are independent — one outcome doesn't change the odds of the next.

Let's derive the formula rather than memorise it. Suppose you want exactly k successes out of n trials. Pick any one specific arrangement — say the first k trials succeed and the remaining n - k fail. Because trials are independent, the probability of that exact sequence is found by multiplying:

P(one specific sequence with k successes) = p × p × ... × p (k times) × q × q × ... × q (n-k times)
                                            = p^k × q^(n-k)

But that's only one arrangement out of many that also give exactly k successes — the successes could land on any k of the n trial-slots. The number of ways to choose which k slots succeed is the combination C(n, k) = n! / (k!(n-k)!). Since these arrangements are mutually exclusive (a trial sequence can't match two different success-patterns at once), we add their probabilities — which, since each has the same probability p^k q^(n-k), just means multiplying by the count:

P(X = k) = C(n, k) × p^k × q^(n-k),   k = 0, 1, 2, ..., n

This is the Binomial PMF. Every piece has a concrete meaning: C(n,k) counts the arrangements, p^k is the chance the chosen k trials all succeed, and q^(n-k) is the chance the rest all fail.

Worked Example — Sixes in an Over

Suppose a hard-hitting batter has, based on this season's data, roughly a 20% chance of hitting a six off any given ball he faces (assume ball-to-ball independence for this model — a simplification, but a standard one). In a 6-ball over, what's the probability he hits exactly 2 sixes?

Here n = 6, p = 0.2, q = 0.8, k = 2.

C(6, 2) = 6! / (2! × 4!) = 15
P(X = 2) = 15 × (0.2)^2 × (0.8)^4
         = 15 × 0.04 × 0.4096
         = 0.24576 ≈ 0.2458

Let's verify this with code and trace it by hand:

from math import comb

n, p, k = 6, 0.2, 2
prob = comb(n, k) * (p**k) * ((1 - p)**(n - k))
print(round(prob, 4))

Tracing it: comb(6, 2) evaluates to 15; p**k is 0.2**2 = 0.04; (1-p)**(n-k) is 0.8**4 = 0.4096; multiplying, 15 × 0.04 × 0.4096 = 0.24576, which round(..., 4) gives as 0.2458. So there's roughly a 24.6% chance of exactly 2 sixes in the over — notably, that is not the same as the intuitive-sounding "20% of 6 balls is about 1.2 sixes on average" — the Binomial gives you the full spread around that average, not a single expected outcome.

Mean and Variance of a Binomial — Derived, Not Assumed

Every textbook states Mean = np and Variance = npq, but few derive them cleanly. Here's the elegant way, using indicator random variables — a trick you'll reuse constantly in JEE-level probability and later in statistics.

Let X_i = 1 if trial i is a success and X_i = 0 otherwise, for i = 1, 2, ..., n. Then the total number of successes is simply X = X_1 + X_2 + ... + X_n. Expectation is linear — the expectation of a sum is the sum of expectations, even if the variables aren't independent — so:

E[X] = E[X_1] + E[X_2] + ... + E[X_n]

Each X_i is 1 with probability p and 0 with probability q, so E[X_i] = 1×p + 0×q = p. Summing n identical terms:

E[X] = np

For variance, we do need independence (which the Binomial's third condition guarantees), because variance of a sum of independent variables is the sum of variances:

Var(X) = Var(X_1) + Var(X_2) + ... + Var(X_n)

To find Var(X_i), use Var(X_i) = E[X_i^2] - (E[X_i])^2. Since X_i only ever takes the values 0 or 1, X_i^2 = X_i always (0² = 0, 1² = 1) — so E[X_i^2] = E[X_i] = p. Then:

Var(X_i) = p - p^2 = p(1 - p) = pq

Summing n identical terms gives the full result:

Var(X) = npq,   Standard deviation = √(npq)

For our sixes example: mean = 6 × 0.2 = 1.2 sixes per over, variance = 6 × 0.2 × 0.8 = 0.96, standard deviation ≈ 0.98. So the batter typically hits somewhere around 0 to 2 sixes an over, with 1.2 as the long-run average across many, many overs — never a guaranteed outcome on any single over.

Misconception Check — "The Average Means Every Trial Looks Average"

A very common error: since p = 0.2, students assume "1 in every 5 balls should be a six" as a near-guarantee, and get confused when an over passes with zero sixes, or with three. The mean np is a long-run average across many repetitions of the whole 6-ball experiment — it says nothing about what any single over must look like. The variance term npq exists precisely to quantify how much real outcomes swing around that average; a mean without a variance is an incomplete description of randomness, and this is exactly why both numbers are computed together, not just the mean.

The Poisson Distribution: When Trials Multiply and Success Gets Rare

The Binomial formula needs a definite n and a non-tiny p. But what happens to counting problems like "number of typing errors on a page," "number of ISRO ground-station signal dropouts in an hour," or "number of failed UPI transactions per minute across a huge user base"? Here there's no clean fixed n of "trials" — every millisecond is technically a chance for something to fail, so n is enormous and p per "trial" is minuscule. What stays meaningful is the average rate, denoted λ (lambda) — for instance, "on average 4 transaction failures occur per minute."

The Poisson distribution is what the Binomial formula turns into as you push n → ∞ and p → 0 while holding their product np = λ fixed. This isn't a separate assumption bolted on — it's a genuine limit, and deriving it is one of the most satisfying algebraic arguments in this whole topic.

Start from the Binomial PMF with p replaced by λ/n:

P(X = k) = C(n, k) × (λ/n)^k × (1 - λ/n)^(n-k)

Expand C(n,k) = n(n-1)(n-2)...(n-k+1) / k!. As n grows very large while k stays fixed, the product n(n-1)...(n-k+1) is a product of k terms each extremely close to n, so it approaches n^k. Substituting:

P(X = k) ≈ [n^k / k!] × (λ^k / n^k) × (1 - λ/n)^(n-k)
         = (λ^k / k!) × (1 - λ/n)^(n-k)

Now split (1 - λ/n)^(n-k) = (1 - λ/n)^n × (1 - λ/n)^(-k). From the definition of Euler's number, (1 - λ/n)^n → e^(-λ) as n → ∞ (this is the same limit that defines e, just with -λ in place of 1). And since k is fixed while n → ∞, the factor (1 - λ/n)^(-k) → 1. Putting it together:

P(X = k) → (λ^k / k!) × e^(-λ) = e^(-λ) × λ^k / k!

That's the Poisson PMF. It requires only one parameter, λ, the average rate of events, and it applies whenever events happen independently, at a roughly constant average rate, and (in principle) one at a time rather than in clumps.

Since Poisson is the limit of Binomial as p → 0, its mean and variance follow directly from the Binomial results: Mean = np → λ, and Variance = npq = np(1-p) → λ × 1 = λ (because p → 0 makes q → 1). A distinctive Poisson fact worth remembering for exams: mean equals variance, both equal to λ. No other named distribution in this chapter has that property.

Worked Example — Cancellations per Hour

An Indian Railways helpline reports that berth cancellations on a particular route come in at an average rate of λ = 4 per hour, spread fairly evenly and independently through the day. What's the probability of getting exactly 2 cancellation calls in the next hour?

P(X = 2) = e^(-4) × 4^2 / 2!
         = 0.018316 × 16 / 2
         = 0.018316 × 8
         = 0.14653 ≈ 0.1465

Checked in code:

from math import exp, factorial

lam, k = 4, 2
prob = exp(-lam) * lam**k / factorial(k)
print(round(prob, 4))

Tracing it: exp(-4) ≈ 0.0183156; lam**k = 4**2 = 16; factorial(2) = 2; so prob = 0.0183156 × 16 / 2 = 0.0183156 × 8 = 0.1465248, and round(..., 4) prints 0.1465. About a 14.7% chance of exactly two cancellations in a given hour — notice we needed no "number of possible callers" at all, only the rate.

Misconception Check — Poisson Needs True Independence

Students often reach for Poisson for anything that looks like "a count of rare events," including situations where it quietly fails. Example: the number of WhatsApp messages in a large family group during a viral forward. This looks countable and "rare-ish" per second, but it is not Poisson, because messages are not independent — one message ("Did you see this?") reliably triggers several replies in a burst. Real independence is a load-bearing assumption, not a formality: violate it and the Poisson formula will systematically underestimate the chance of large bursts and overestimate the chance of a quiet, evenly-spread hour. Always ask "does one event happening make a nearby event more or less likely?" before reaching for Poisson.

The Normal Distribution: The Shape Randomness Converges To

Return to the CBSE result-day marks list. Why does it come out bell-shaped, almost regardless of the subject or the exam? The deep reason is the Central Limit Theorem: whenever a measured outcome is effectively the sum (or average) of many small, roughly independent contributing factors, the distribution of that outcome tends toward the same bell shape, no matter what the individual factors' own distributions looked like. A student's raw score is influenced by dozens of small, semi-independent factors — familiarity with each topic, luck of question phrasing, alertness on exam day — and their combined effect washes out into the Normal shape.

There's a second, more direct route to the same curve: take the Binomial distribution and let n grow large while p stays fixed (not shrinking to 0, unlike the Poisson limit). The discrete bars of the Binomial PMF get more numerous and less jagged, and their outline converges to a smooth bell curve — this is the De Moivre–Laplace theorem, historically the first appearance of the Normal distribution, discovered as an approximation to the Binomial almost a century before Gauss's name got attached to it. You can see this progression directly in the diagram below: the symmetric Binomial bars in the first panel are already hinting at the smooth curve in the third.

The Normal PDF, with mean μ and standard deviation σ, is:

f(x) = [1 / (σ√(2π))] × e^(-(x-μ)² / (2σ²))

You won't be asked to derive this formula from the Central Limit Theorem at this stage — that derivation genuinely needs university-level analysis. What you should understand precisely is what each symbol controls: μ slides the whole curve left or right (it's the peak's location and, by symmetry, both the mean and the median); σ stretches or squeezes the curve horizontally (a small σ gives a tall, narrow spike — most values close to μ; a large σ gives a short, wide spread). The total area under the curve is exactly 1, because it must account for all possible outcomes — proving that fact rigorously needs a calculus result (the Gaussian integral) that is beyond this chapter, but the interpretation — area equals probability — is what you'll use constantly.

The Empirical Rule and Standardization

To compare across different means and spreads, we standardize: convert any Normal value x into a z-score, which measures how many standard deviations away from the mean it sits:

z = (x - μ) / σ

Every Normal distribution, once standardized, obeys the same universal proportions, known as the 68-95-99.7 empirical rule:

• About 68% of values fall within ±1σ of the mean (-1 ≤ z ≤ 1)
• About 95% fall within ±2σ
• About 99.7% fall within ±3σ

These aren't arbitrary — they come from integrating the Normal PDF over those intervals (a calculation done once, tabulated forever after as the "standard normal table" you'll use formally in Class 11–12 statistics and in JEE Main/BITSAT probability sections).

Worked Example — JEE Main Percentile

Suppose scores on a mock JEE Main test are modelled as Normal with μ = 150 and σ = 30 (out of 300). What fraction of students score above 210?

z = (210 - 150) / 30 = 60 / 30 = 2

A score of 210 sits exactly 2 standard deviations above the mean. By the empirical rule, 95% of students fall within ±2σ, so the remaining 5% falls outside that range — split evenly between the two tails (by symmetry of the bell curve), giving 2.5% in each tail. So roughly 2.5% of students score above 210 — that's the top 2.5% of the cohort, a genuinely useful number for estimating a percentile cutoff without needing a full z-table lookup.

Misconception Check — "What's the Probability of Scoring Exactly 210?"

This question, asked of a continuous Normal model, has answer zero — and that surprises most students the first time they meet it. Here's why it's not a trick: there are infinitely many possible scores between, say, 209 and 211 (209.001, 209.0001, and so on without end), so the "share" of probability assigned to any single infinitely-precise point must shrink to nothing. Probability under a continuous distribution only makes sense over an interval — "between 205 and 215" is a real, computable, nonzero probability (it's the area of the curve between those two x-values); "exactly 210.000..." is not. This is the single most important discrete-vs-continuous distinction in the whole chapter, and exam-setters love testing it directly.

Seeing All Three Together

The diagram below plots all three distributions from this chapter using their actual computed values — not sketches. Panel 1 shows the Binomial(n=10, p=0.5) bars: symmetric, discrete, peaking at k=5. Panel 2 shows Poisson(λ=3): discrete but skewed right, with a longer tail toward larger counts — a direct visual signature of rare-event distributions. Panel 3 shows the standard Normal curve, continuous and smooth, with the middle 68% (within ±1σ) shaded to make the empirical rule concrete.

Choosing the Right Distribution

Under exam pressure, the fastest way to pick correctly is to ask a short sequence of questions about the situation, not about the numbers:

• Is the outcome a count of successes out of a fixed, known number of independent tries, each with the same success chance? → Binomial. (Fixed n is the giveaway: "in 10 free throws," "out of 20 MCQs," "in a batch of 50 chips.")
• Is the outcome a count of independent events over a continuous stretch of time or space, with a known average rate but no natural "number of trials"? → Poisson. (No fixed n is the giveaway: "per hour," "per kilometre of road," "per page.")
• Is the outcome a continuous measurement that results from adding up many small independent effects? → Normal. (Look for "measurement," "score," "height," "error," rather than "count.")

One more connecting fact worth remembering for JEE/BITSAT-level questions: when n is large and p isn't tiny, Binomial itself starts to look Normal (De Moivre–Laplace) — so in that regime, either model gives you a good approximate answer, and you'll sometimes be explicitly asked to use the Normal approximation to sidestep computing an enormous factorial.

Practice — Active Recall

Work these without looking back at the worked examples, then check your reasoning against the methods above.

1. A fair coin is tossed 8 times. Using the Binomial PMF, find P(exactly 5 heads). (Hint: n=8, p=0.5, k=5.)
2. Find the mean and variance of the number of heads in Question 1, using the np and npq formulas derived above.
3. A call centre receives complaints at an average rate of λ = 5 per hour, independently. Find P(exactly 3 complaints in the next hour).
4. Explain in one or two sentences why the number of monsoon-season potholes reported per kilometre of a highway is a better candidate for a Poisson model than for a Binomial model.
5. A machine fills juice bottles with volumes Normal(μ=200 ml, σ=5 ml). Using the empirical rule, estimate the percentage of bottles filled below 190 ml.
6. True or false, with a one-line justification: "For a continuous random variable, P(X = μ) is the highest single probability value in the distribution." (Careful — this tests the exact-point misconception from this chapter.)

Rough answer check: (1) C(8,5)(0.5)^8 = 56/256 ≈ 0.2188. (2) mean = 4, variance = 2. (3) e^-5 × 5^3/3! ≈ 0.1404. (4) potholes have no natural fixed "number of trials" per kilometre — only a rate. (5) 190 ml is 2σ below the mean, so about 2.5% of bottles (using the same tail logic as the JEE example). (6) False — P(X=μ) is exactly 0 for any continuous variable; the correct statement involves the density being highest at μ, not the probability.

Summary

A Binomial distribution counts successes across a fixed number of independent, identical trials, with PMF C(n,k)p^kq^(n-k), mean np, and variance npq — both derived cleanly from indicator variables rather than assumed. A Poisson distribution is what the Binomial becomes in the limit of enormous n and vanishing p with np = λ held fixed, giving PMF e^(-λ)λ^k/k!, with the distinctive property that mean and variance are both exactly λ. A Normal distribution describes continuous measurements built from many small independent effects (via the Central Limit Theorem), or equivalently the smoothed-out limit of Binomial bars when n is large and p isn't tiny (De Moivre–Laplace); its shape is fixed by μ and σ, standardized scores follow the universal 68-95-99.7 rule, and probability only ever means area under the curve over an interval — never the height at one exact point. Recognising which of the three questions you're actually being asked — fixed trials, rate over a continuum, or a continuous sum-of-effects measurement — is most of the battle, both on CBSE boards and in JEE/BITSAT-level probability problems.