90% Accurate, Still Usually Wrong Probability

Toss three fair coins together, giving 8 equally likely outcomes, and ask for the probability of getting at least two heads: exactly 4 of those 8 outcomes qualify, so the answer is 1/2. But now suppose someone tells you the first coin specifically landed tails -- does that change anything? It changes everything: restricted to just the 4 outcomes where the first coin is tails, only ONE of them (tails-heads-heads) still has at least two heads overall, shrinking the probability all the way down to 1/4. The extra information didn't just refine the answer slightly -- it replaced the entire sample space being measured against, from all 8 outcomes down to just the 4 consistent with what's now known.

That shrinking-sample-space idea becomes a single formula: the conditional probability of E given F, written P(E|F), equals P(E and F) divided by P(F) -- provided F itself has a nonzero chance of happening at all. Three clean properties follow directly: the whole original sample space, conditioned on F, still has probability exactly 1 (nothing is lost, just rescaled); the addition rule for 'or' still works inside the condition, P((A or B)|F) = P(A|F)+P(B|F)-P((A and B)|F); and the complement rule survives too, P(not-E|F) = 1-P(E|F). None of ordinary probability's structure breaks once you condition on new information -- it just all operates inside the smaller, updated space.

Rearranging the conditional-probability formula slightly gives an equally useful multiplication rule: P(E and F) = P(E) x P(F|E), and just as validly, P(E and F) = P(F) x P(E|F). This turns conditional probability into a construction tool rather than just an analysis tool -- drawing 2 kings and then an ace from a 52-card deck, without replacement, becomes a simple chain: 4/52 for the first king, then 3/51 for the second (one king is already gone), then 4/50 for the ace (2 cards gone, but all 4 aces remain), multiplying out to a clean 2/5525. The rule extends cleanly to three, four, or any number of events chained together, each step conditioned on everything that came immediately before it.

Two events are called independent precisely when knowing one occurred changes nothing about the other's probability -- formally, P(E and F) = P(E) x P(F), which is equivalent to P(E|F) simply equalling the plain P(E). It's tempting to confuse this with 'mutually exclusive' (events that can never both happen), but they are not remotely the same idea: two mutually exclusive events, as long as each has a nonzero probability on its own, can NEVER be independent -- because knowing one happened tells you the other one absolutely did NOT, which is about as much new information as it's possible to get. Drawing a spade and drawing an ace from one card, by contrast, ARE independent (each suit contains exactly one ace, keeping the proportion identical whether or not you already know the suit) -- a genuinely different, and genuinely compatible, relationship.

A partition of a sample space is a set of events that are pairwise disjoint (no two overlap) and collectively exhaustive (together, they cover every single outcome). Any event and its own complement automatically form the simplest possible partition, but a partition can just as easily have many pieces -- like splitting 'which bag a ball came from' into 'Bag I' and 'Bag II,' or 'which machine produced this item' into 'Machine A,' 'Machine B,' and 'Machine C.' Given such a partition, the theorem of total probability lets any other event A's overall probability be rebuilt piece by piece: P(A) equals the sum, across every piece Ei of the partition, of P(Ei) x P(A|Ei) -- in other words, weight each piece's own conditional contribution by how likely that piece is in the first place, then add all the weighted contributions together.

Everything so far computes probabilities running forward in time -- from a cause (which bag, which machine) to an effect (red ball, defective item). Bayes' theorem runs the exact same machinery backwards: given that the effect A has already been observed, it recovers the probability that any specific piece Ei of the partition was the actual cause, via P(Ei|A) equal to P(Ei) x P(A|Ei), divided by the same sum used in the total probability theorem. The payoff is genuinely striking: a lab test that correctly flags a disease 90% of the time, with only a 1% false-positive rate among healthy people, sounds about as reliable as a test gets -- yet if only 0.1% of the wider population actually has the disease, a random positive result still comes out only about 8.3% likely to be a genuine case. The test's own accuracy and a specific result's trustworthiness are two completely different numbers, and mixing them up is exactly the mistake Bayes' theorem exists to correct.

The Reverend Thomas Bayes worked out this reverse-probability formula himself, but it wasn't published until 1763 -- after his death, when a colleague brought the work to the Royal Society on his behalf. It joins a chapter of genuinely old, occasionally strange probability history: Girolamo Cardano's own gambling treatise was published posthumously too, exactly a century earlier; Galileo, asked to explain a specific dice puzzle, correctly worked out that three dice landing on a total of 10 is genuinely more likely than landing on 9, simply because more of the underlying triples of numbers add up to 10 than to 9; and Jacob Bernoulli's landmark book on the subject, containing the discovery of what's now called the Binomial distribution, was ALSO published only after Bernoulli's own death, in 1713. A surprising number of probability theory's foundational results reached print only once their discoverers were no longer around to see it.

This chapter's own opening paragraph still describes three topics as coming up: conditional probability and Bayes' theorem (which do follow), then 'an important concept of random variable and its probability distribution,' the mean and variance of that distribution, and finally 'an important discrete probability distribution called Binomial distribution.' None of that last set actually appears anywhere in the current chapter -- it was trimmed out in NCERT's 2022 rationalisation, along with the two worked examples that used to illustrate it, and the chapter now ends, right after Bayes' theorem and a Miscellaneous Exercise, with a summary and a historical note instead. One question in that very Miscellaneous Exercise still asks for the probability that 'at most 6 of a random sample of 10' right-handed-or-not people turn out right-handed -- a question that technically still needs exactly the Binomial-distribution tool the current book no longer teaches anywhere, a small leftover fossil of the fuller chapter this one used to be.