8 topics. Each one opens with its key idea, then the detail you need to use it.
1. The Law of Large Numbers and simulation
Key idea: Probability is a long-run proportion, and the short run has no duty to balance out.
Probability describes what happens over many repetitions of a random process. Suppose a basketball player makes 70% of her free throws. The Law of Large Numbers says that her running proportion of makes gets closer to 0.70 as her number of shots grows, which is what convergence means here. In 10 shots she could easily make 5 or 9, but in 1,000 shots her proportion will almost surely sit near 0.70. The law says nothing about the short run. After five misses in a row, her next shot is still a 0.70 chance, because a random process has no memory. Believing that a make is "due" is the gambler's fallacy, and the exam tests it directly.
Simulation uses the same idea to estimate probabilities that are hard to calculate. A complete simulation has four parts. First, assign random digits to outcomes so the probabilities match: for 0.70, digits 0 to 6 mean a make and 7 to 9 mean a miss. Second, describe one trial and what you record, such as "read five digits and count the makes." Third, repeat the trial many times. Fourth, report the proportion of trials in which the event happened. If the real process samples without replacement, say that you will skip repeated labels. More trials give an estimate that varies less from run to run, which is the Law of Large Numbers again.
2. Probability rules: the complement and the addition rule
Key idea: "Or" means add, then subtract the overlap; "at least one" means 1 minus "none."
Every probability lies between 0 and 1, and the probabilities of all the possible outcomes add to exactly 1. So an answer like 3.5 or 35 can never be a probability, and getting one signals an arithmetic or unit slip. The complement rule follows from that total of 1: P(not A) = 1 − P(A). It is the quickest route through any "at least one" question, because "at least one" is the complement of "none."
The addition rule gives the chance that A or B happens: P(A or B) = P(A) + P(B) − P(A and B). You subtract the overlap because adding P(A) and P(B) counts the people in both groups twice. Suppose 45% of students at a school play soccer, 30% play basketball and 12% play both. Then P(soccer or basketball) = 0.45 + 0.30 − 0.12 = 0.63. In a Venn diagram the 12% sits in the shared middle, and you want it counted once, not twice.
Mutually exclusive events cannot happen together, so P(A and B) = 0 and the rule shrinks to P(A) + P(B). A single card cannot be both a heart and a spade, so those events are mutually exclusive. Hearts and kings overlap in the king of hearts, so P(heart or king) = 13/52 + 4/52 − 1/52 = 16/52.
3. Two-way tables, conditional probability and Simpson's paradox
Key idea: The event after the bar in P(A | B) tells you the denominator.
A two-way table makes conditional probability concrete. Imagine 200 students: 120 juniors, of whom 30 drive to school, and 80 seniors, of whom 48 drive. In total, 78 students drive, so P(drives) = 78/200 = 0.39. P(drives | senior) means "among seniors only," so the denominator is 80, giving 48/80 = 0.60. P(senior | drives) flips the condition, so now the denominator is the 78 drivers, giving 48/78 ≈ 0.615. The two answers differ because they divide by different groups. The formula P(A | B) = P(A and B) / P(B) says the same thing in symbols: find the joint share, then divide by the share of the given group.
A segmented bar chart shows conditional distributions as pictures. Each bar is one group scaled to 100%, so you compare shares within groups rather than raw counts, which matters when the groups have different sizes.
Simpson's paradox happens when a comparison that holds inside every group reverses once the groups are combined. Say Hospital A saves 60% of its severe cases (240 of 400) and 95% of its mild cases (95 of 100). Hospital B saves 55% of severe cases (55 of 100) and 92% of mild cases (368 of 400). A is better in both groups. Yet overall, A saves 335 of 500 patients (67%) while B saves 423 of 500 (84.6%). The reversal happens because A treats mostly severe cases, which have lower survival everywhere. Severity is a lurking variable, and the combined table hides it.
4. Independence, the multiplication rule and drawing without replacement
Key idea: Test independence by comparing P(A | B) with P(A), and multiply plain probabilities only when that check passes.
Two events are independent when knowing that one happened does not change the chance of the other: P(A | B) = P(A). To check independence with a table, compare a conditional probability with the overall one. In the driving table, P(drives) = 0.39 but P(drives | senior) = 0.60, so driving and grade level are not independent events.
The general multiplication rule is P(A and B) = P(A) × P(B | A). For independent events, P(B | A) is just P(B), so you multiply the two probabilities directly. If each of 4 free throws is an independent 0.7 chance, P(all 4 made) = 0.7⁴ ≈ 0.240, so P(at least one miss) ≈ 0.760.
Drawing without replacement breaks independence, because each draw changes what is left in the bag. From a bag of 5 red and 7 blue marbles, P(two reds) = (5/12)(4/11) = 20/132 ≈ 0.152. When the population is large, one draw barely changes it, so the 10% condition lets you treat draws as independent if the sample is no more than 10% of the population.
Tree diagrams chain these steps together. Suppose 2% of people carry a condition, a test flags 95% of carriers, and it wrongly flags 8% of non-carriers. Then P(positive) = (0.02)(0.95) + (0.98)(0.08) = 0.0974, and P(carrier | positive) = 0.019 / 0.0974 ≈ 0.195. Most positives are false alarms because carriers are so rare.
Independent does not mean mutually exclusive. If A and B each have some chance of happening but cannot happen together, learning that A happened makes B impossible, so the events are dependent.
5. Discrete random variables, expected value and transformations
Key idea: Expected value is a long-run average per trial, and variances add, never standard deviations.
A discrete random variable takes a countable list of numeric values, each with a probability. The probabilities must each lie between 0 and 1 and add to exactly 1. The expected value is a weighted average: μ = Σ x·P(x). Take a $3 carnival game that pays $10 with probability 0.10, $4 with probability 0.25 and nothing with probability 0.65. E(winnings) = 10(0.10) + 4(0.25) + 0(0.65) = $2, so the expected net gain is −$1 per play. No single play loses exactly $1. The −$1 is the average result per play over many plays.
Spread comes from the variance, σ² = Σ (x − μ)² P(x). Here it is 8²(0.10) + 2²(0.25) + (−2)²(0.65) = 10, so σ = √10 ≈ $3.16.
A linear transformation Y = a + bX turns the mean into a + bμ and the standard deviation into |b|σ. Adding a slides every value the same distance, so the spread does not change, and neither does the shape. Daily highs averaging 20°C with SD 5°C become 32 + 1.8(20) = 68°F with SD 1.8(5) = 9°F.
When you combine random variables, means add or subtract directly. For spread, use variances, and only when X and Y are independent: Var(X ± Y) = Var(X) + Var(Y). The plus sign holds for differences too. If X has mean 30 and SD 4 and an independent Y has mean 20 and SD 3, then X − Y has mean 10 and SD √(16 + 9) = 5, not 7 and not 1.
6. The binomial distribution and the normal approximation
Key idea: Check BINS before using the binomial formula, and check both np and n(1 − p) before using the normal approximation.
A binomial random variable counts successes in a fixed number of trials. Check four conditions, often remembered as BINS. Binary: each trial is a success or a failure. Independent: one trial does not affect another. Number: n is fixed in advance. Same: every trial has the same success probability p. Counting trials until the first success is a geometric setting instead, because n is not fixed.
The binomial formula is P(X = k) = C(n, k) p^k (1 − p)^(n − k). The C(n, k) term counts the different orders in which the k successes can appear. A student guesses on 10 multiple-choice questions with 4 choices each, so p = 0.25. P(exactly 3 right) = C(10, 3)(0.25)³(0.75)⁷ = 120 × 0.015625 × 0.1335 ≈ 0.250. P(at least 1 right) = 1 − 0.75¹⁰ ≈ 1 − 0.056 = 0.944. The mean of a binomial distribution is np = 2.5 and the SD is √(np(1 − p)) = √1.875 ≈ 1.37.
When n is large, a binomial histogram looks bell-shaped, so a normal model gives a good approximation. The condition is np ≥ 10 and n(1 − p) ≥ 10, and both parts must hold. With n = 400 and p = 0.30, np = 120 and n(1 − p) = 280, so you can use a normal model with mean 120 and SD √84 ≈ 9.17. With n = 150 and p = 0.04, np = 6, so a normal model would be poor even though n(1 − p) = 144 is large.
7. Density curves, the normal distribution and the empirical rule
Key idea: Sketch the curve, mark μ ± 1σ, 2σ and 3σ, and split every percentage by symmetry before you answer.
A density curve models a continuous variable. It never dips below the horizontal axis, and the total area under it equals 1. The probability that X falls in an interval equals the area above that interval, so if the area between 10 and 20 is 0.35, the probability is 0.35. Because a single point has no width, P(X = 15) = 0 for a continuous variable, which makes P(X < 15) and P(X ≤ 15) equal.
A uniform distribution is the simplest density curve: a flat rectangle. If a bus wait is uniform from 0 to 12 minutes, the height must be 1/12 so that the total area is 1. P(wait between 3 and 7 minutes) = 4 × 1/12 = 1/3.
A normal distribution is symmetric and bell-shaped, with its mean equal to its median. The mean μ sets the centre and the standard deviation σ sets the width. The empirical rule, also called the 68-95-99.7 rule, gives quick areas: about 68% of values lie within 1σ of μ, 95% within 2σ, and 99.7% within 3σ. Within 1σ is 68%, so each half is 34%, and each tail beyond 1σ holds 16%. Beyond 2σ on one side is 2.5%, and beyond 3σ on one side is 0.15%.
Try battery life with μ = 40 hours and σ = 4. About 68% of batteries last 36 to 44 hours, and about 2.5% last more than 48 hours. The share lasting 32 to 44 hours is 47.5% + 34% = 81.5%.
8. Z-scores, percentiles and the z table
Key idea: The table always gives the area to the left; decide which side you need before you look anything up.
A z-score tells you how many standard deviations a value sits from the mean: z = (x − μ)/σ. A positive z means above the mean and a negative z means below it. Because z has no units, it lets you compare values on different scales. A biology score of 82 on a test with mean 70 and SD 8 has z = 1.5. A chemistry score of 88 on a test with mean 73 and SD 12 has z = 1.25. The biology score is the stronger result, even though it is lower in raw points.
For a normal distribution, a z-score turns into an area. A standard normal z table gives the area to the LEFT of z. On a calculator, normalcdf(lower, upper) gives the area between two bounds, so for a left area you use a very large negative lower bound, such as normalcdf(−1000, z). That left area is the percentile, the percent of values at or below x. Suppose nightly sleep is normal with mean 7.2 hours and SD 0.8. For 6 hours, z = (6 − 7.2)/0.8 = −1.5, and the table gives 0.0668, so 6 hours is about the 7th percentile. A negative z always lands below the 50th percentile.
For a right tail, subtract the table value from 1. For 8.4 hours, z = 1.5, so P(more than 8.4) = 1 − 0.9332 = 0.0668. For an interval, subtract two left areas: P(6.4 to 8.4) = 0.9332 − 0.1587 = 0.7745.
Working backward turns a percentile into a value. The 90th percentile has z ≈ 1.28, which is what invNorm(0.90) returns. So x = μ + zσ = 7.2 + 1.28(0.8) ≈ 8.22 hours.