Topic 6: Probability Dr J Frost (jfrost@tiffin.kingston.sch.uk) Last modified: 18th July 2013 Slide guidance Key to question types: SMC Senior Maths Challenge Uni Questions used in university interviews (possibly Oxbridge). www.ukmt.org.uk The level, 1 being the easiest, 5 the hardest, will be indicated. BMO British Maths Olympiad Those with high scores in the SMC qualify for the BMO Round 1. The top hundred students from this go through to BMO Round 2. Questions in these slides will have their round indicated. MAT Maths Aptitude Test Admissions test for those applying for Maths and/or Computer Science at Oxford University. University Interview Frost A Frosty Special Questions from the deep dark recesses of my head. Classic Classic Well known problems in maths. STEP STEP Exam Exam used as a condition for offers to universities such as Cambridge and Bath. Slide guidance ? ο Any box with a ? can be clicked to reveal the answer (this works particularly well with interactive whiteboards!). Make sure you’re viewing the slides in slideshow mode. For multiple choice questions (e.g. SMC), click your choice to reveal the answer (try below!) Question: The capital of Spain is: A: London ο» B: Paris ο» C: Madrid οΌ Topic 6: Probability Part 1 – Manipulating Probabilities Part 2 – Random Variables a. b. c. d. e. Random Variables Discrete and Continuous Distributions Mean and Expected Value Uniform Distributions Standard Deviation and Variance Part 3 – Common Distributions a. b. c. d. e. Binomial Bernoulli Poisson Geometric Normal/Gaussian Some starting notes Only some of those reading this will have done a Statistics module at A Level. Therefore only GCSE knowledge is assumed. There is some overlap with the field of Combinatorics. For probability problems relating to ‘arrangements’ of things, look there instead. Probability and Stats questions in… In university interviews… In SMC In STEP Probability questions frequently come up (although not technically requiring any more than GCSE theory). In my experience, applicants tend to do particularly bad at these questions. Harder probability questions are quite rare (although one appeared towards the end of 2012’s paper) Two questions at the end of every paper. You could avoid these, but you broaden your choice if you prepare for these. In BMO Used to be moderately common, but less so nowadays But some basic probability/statistics will broaden your maths ‘general knowledge’. You’ll know for example what scientists at CERN mean when they refer in the news to the “5π test” needed to verify that a new particle has been discovered! Topic 6 – Probability Part 1: Manipulating Probabilities Events and Sets An event is a set of outcomes. “Even number thrown on a die” = {2, 4,? 6} Given that events can be represented as sets, we can use set operations. Suppose πΈ = {2, 4, 6}, say the event of throwing an even number, and π = {2, 3, 5}, say the event of throwing a prime number. Then: πΈ ∩ π = {2}? ? 5, 6} πΈ ∪ π = {2, 3, 4, ∩ means set intersection. It gives the items which are members of both sets. It represents “numbers that are even AND prime”. ∪ means set union. It gives the items which are members of either. It represents “numbers that are even OR prime”. GCSE Recap When π΄ and π΅ are mutually exclusive (i.e. π΄ and π΅ can’t happen at the same time, or more formally π΄ ∩ π΅ = ∅, where ∅ is the ‘empty set’)… π π΄ ∪ π΅ = π π΄ +?π(π΅) When π΄ and π΅ are independent (i.e. π΄ and π΅ don’t influence each other)… π π΄ ∩ π΅ = π π΄ ×? π(π΅) When π΄ and π΅ are mutually exclusive… π π΄∩π΅ =0 ? More useful identities When π΄ and π΅ are not mutually exclusive… π π΄ ∪ π΅ = π π΄ + π π΅ ?− π(π΄ ∩ π΅) When π΄ and π΅ are independent and mutually exclusive… π π΄ ∪ π΅ = π π΄ + π π΅ ?− π π΄ π(π΅) Conditional probabilities We might want to express the probability of an event given that another occurred: The probability that A occurred given B occurred. π π΄|π΅ To appreciate conditional probabilities, consider a probability tree: 1st pick 2nd pick 3 6 4 7 Red 3 7 Green 3 6 Red Green This represents the probability that a green counter was picked second GIVEN that a red counter was picked first. Conditional probabilities Using the tree, we can construct the following identity for condition probabilities: π(π΅|π΄) π(π΄) B A π π΄ ∩ π΅ = π π΄ π(π΅|π΄) or π π΄∩π΅ π π΅π΄ = π π΄ π(π΄ ∩ π΅) Conditional probabilities If π΄ and π΅ are independent, then what is π(π΄|π΅)? The Common Sense Method If π΄ and π΅ are independent, then the probability of π΄ occurring is not affected by whether π΅ occurred, so: π π΄π΅ = ? π(π΄) The Formal Method π π΄∩π΅ π π΄π΅ = π π΅ = π π΄ π π΅ π π΅ ? Because π΄ and π΅ are independent. Examples 1 2 The events A and B are independent with π π΄ = 4 and π π΄ ∪ π΅ = 3 . Find: (Source: Edexcel) π π΅ π(π΅′ |π΄) π π΄∪π΅ =π π΄ +π π΅ −π π΄ π π΅ 2 1 1 = + π π΅ − π? π΅ 3 4 4 5 π π΅ = 9 π π΅′ π΄ = π(π΅′ ) because π΄ and π΅ are independent. π π΄′ ∩ π΅ π π΄′ ∩ π΅ = π π΄′ × π π΅ 3 5 15 5 = 4 × 9 = 36 = 12 ? 4 So π π΅′ π΄ = 9 ? Bayes’ Rule Bayes’ Rule relates causes and effects. It allows us find the probability of the cause given the effect, if we know the probability of the effect given the cause. π πΈπΆ π πΆ π πΆπΈ = π πΈ Dr House is trying to find the cause of a disease. He suspects Lupus (as he always does) due to their kidney failure. The probability that someone has this symptom if they did have Lupus is 0.2. The probability that a random patient has kidney damage is 0.001, and the probability they have Lupus 0.0001. What is the probability they have Lupus given their observed symptom? 0.2 × 0.0001 π πΏπΎ = = 0.002 ? 0.001 Bayes’ Rule But we don’t always need to know the probability of the effect. π· π¬πͺ π· πͺ π· πͺπ¬ = π· π¬ Notice that in the distribution π(πΆ|πΈ), πΈ is fixed, and the distribution is over different causes, where π∈πΆ π πΆ πΈ = 1. This suggests we can write: π· πͺ π¬ =ππ· π¬ πͺ π· πͺ where π is a normalising constant that is set to ensure our probabilities add up to 1, i.e. π πΆ πΈ + π πΆ ′ πΈ = 1 Question: The probability that a game is called off if it’s raining is 0.7. The probability it’s called off if it didn’t rain (e.g. due to player illness) is 0.05. The probability that it rains on any given day is 0.2. Andy Murray’s game is called off. What’s the probability that rain was the cause? Bayes’ Rule Question: The probability that a game is called off if it’s raining is 0.7. The probability it’s called off if it didn’t rain (e.g. due to player illness) is 0.05. The probability that it rains on any given day is 0.2. Andy Murray’s game is called off. What’s the probability that rain was the cause? Write down information: π πΆ π = 0.7 π πΆ π ′ = 0.05 ? So π π ′ = 0.8 π π = 0.2 Then using Bayes’ Rule: π π πΆ = π π πΆ π π π = 0.14π π π ′ πΆ = ππ πΆ π ′ π π ′ = 0.04π ? = 1, so π = But π π πΆ + π π ′ πΆ = 1. So 0.18π Then π· πΉ πͺ = π. ππ × ππ π = π π 50 . 9 Topic 6 – Probability Part 2: Random Variables Random Variables A random variable is a variable which can have multiple values, each with an associated probability. The variable can be thought of as a ‘trial’ or ‘experiment’, representing something which can have a number of outcomes. A random variable has 3 things associated with it: The values the random variable can have (e.g. outcomes of the throw of a die) 1 The outcomes 2 A probability function The probability associated with each outcome. Parameters These are constants used in our probability function that can be set (e.g. number of throws) 3 (formally known as the ‘support vector’) Example random variables This symbol means “for all” Random variable (X) Outcomes Parameters? The single throw of a fair die. {1, 2, 3, 4, 5, 6} None The single throw of an unfair die. {1, 2, 3, 4, 5, 6} We can set the probability of ? each outcome: p1, p2, …, p6 ? ? We use capital letters for random variables. Parameters are values we can control, but do not change across different outcomes. We’ll see plenty more examples. ? Probability Function 1 π π=π₯ = ?6 π π = π₯π = ππ ∀π₯ ∀π ? For an outcome x and a random variable X, we express the probability as π(π = π₯), meaning “the probability that the random variable X has the outcome x”. We sometimes write π(π₯) for short, with a lowercase p. In this example, we can use the probability associated with the particular outcome. We sometimes use π₯π to mean the ith outcome. Sketching the probability function It’s often helpful to show the probability function as a graph. Suppose a random variable X represents the single throw of a biased die: Probabilities P(X) 1 2 3 4 X 5 6 Outcomes Discrete vs Continuous Distributions Discrete distributions are ones where the outcomes are discrete, e.g. throw of a die, number of Heads seen in 10 throws, etc. In contrast continuous distributions allow us to model things like height, weight, etc. Here’s two possible probability functions: Discrete Continuous 0.3 P(X=h) P(X=k) 0.4 0.2 0.1 1.2m 1.4m 1.6m 1.8m 2.0m 2.2m 2.4m 1 2 3 4 Number of times target hit (k) Height of randomly picked person (h) Discrete vs Continuous Distributions Discrete P(X=k) 0.4 Probabilities add up to 1. i.e. π₯ π π₯ = 1 0.3 All probabilities must be between 0 and 1, i.e. 0 ≤ π π₯ ≤ 1 ∀π₯. 0.2 0.1 1 2 3 4 Number of times target hit (k) We call the probability function the: Probability mass function (PMF for short) Because our probability function is ultimately just a plain old function (provided it meets the above properties), we often see the function written as “π(π₯)” rather than π(π₯). Continuous Distributions π.π π.π i.e. We find the area under the graph. Note that the area under the entire graph will be 1: +∞ π π₯ ππ₯ = 1 Does it make sense to talk about the probability of someone being exactly 2m? π ?π π π P(X=h) π π. π ≤ π < π. π = Clearly not, but we could for example find the probability of a height being in a particular range. −∞ 1.0m 1.2m 1.4m 1.6m 1.8m 2.0m 2.2m 2.4m 2.6m Height of randomly picked person (h) The probability associated with a particular value is known as the probability density. It’s value alone is not particular meaningful (and can be greater than 1!), but finding the area in a range gives us a probability mass. This is similar to histograms, where the y-axis is the ‘frequency density’, and finding the area under the bars gives us the frequency. Probability Density Question: Archers fire arrows at a target. The probability of the arrow being a certain distance from the centre of the target is proportional to this distance. No archer is terrible enough that his arrow will be more than 1m from the centre. What’s the probability that an arrow is less than 0.5m from the centre? Source: Frosty Special Answer: π π₯ ≤ 0.5π = 0.25 ? 2 Probability is proportional to distance. P(X=x) Maximum distance is 1m. Since area under graph must be 1, then maximum probability density must be 2, so that the area of triangle is ½ x 2 x 1 = 1. 0.5 1 Distance of arrow from centre (x) We’re finding the probability of the arrow being between 0m and 0.5m, so find the area under the graph in this region. We can see this will be 0.25. Probability Density Question: Archers fire arrows at a target. The probability of the arrow being a certain distance from the centre of the target is proportional to this distance. No archer is terrible enough that his arrow will be more than 1m from the centre. What’s the probability that an arrow is less than 0.5m from the centre? Source: Frosty Special Alternatively, using a cleaner integration approach: Step 1: Use the information to express the proportionality relationship: ? π π₯ ∝π₯ , so π π₯ = ππ₯. Step 2: Determine constant by using the fact that 1 π ?0 π π π π ππ₯ ππ₯ = 2 So 2 = 1 and thus π = 2 Step 3: Finally, integrate desired range. 0.5 0 ? 2π₯ ππ₯ = 0.25 π₯ ππ₯ = 1 Mean and Expected Value Mean of a Sample Mean of a Random Variable But what about the mean of a random variable X? This is known as the “expected value of X”, written E[X]. It can be calculated using: The process of using a random variable to give us some values is known as sampling. For example, we might have measured the heights of a sample of people: πΈπ = or πΈ π = The mean of a sample you’ve known how to do since primary school: π₯= π₯π π Archery scores: 57, 94, 25, 42 57 + 94 + 25 + 42 ? = 54.5 π₯= 4 π₯ π(π₯) π π₯ ππ₯ ∞ π₯ −∞ depending on whether your variable is discrete or continuous. X is “times target hit out of 3 shots”. x 0 1 2 3 P(X=x) 0.25 0.5 0.05 0.2 πΈπ ? + 2 × 0.05 = 0 × 0.25 + 1 × 0.5 Expected Value Question: Two people randomly think of a real number between 0 and 100. What is the expected difference between their numbers? (i.e. the average range) (Source: Frosty Special) (Hint: Make your random variable the difference between the two numbers ) As with many problems, it’s easier to consider a simpler scenario. Consider just say integers between 0 and 10. How many ways can the numbers be chosen if the range is 0? Or the range is 1? Or 2? What do you notice? Step 1: Use the information to express the proportionality relationship: π π₯ ∝ 100 − π₯ We can consider the two numbers (with range π₯), as a ‘window’ which we can ‘slide’ in the 0 to 100 region. The bigger?the window, the less we can slide it. If they were to choose 0 and 100, we can’t slide at all. Step 2: Determine constant by using the fact that π π π π₯ ππ₯ = 1 π π₯ = 100π − ππ₯ π Integrating we get 100ππ₯ − 2 π₯ 2 . If the?limits are 0 and 100, we get 5000π = 1 1, so π = 5000 Expected Value Question: Two people randomly think of a real number between 0 and 100. What is the expected difference between their numbers? (i.e. the average range) (Source: Frosty Special) (Hint: Make your random variable the difference between the two numbers ) Step 3: Finally, given our known PDF, find E[X] 100 πΈπ = π₯ π π₯ ππ₯ 100 0 π₯ π? 100 − π₯ = ππ₯ 0 1 = 33 3 One of the harder problem sheet exercises is to consider what happens when we introduce a 3rd number! Modifying Random Variables We often modify the value of random variables. Example: X = outcome of a single throw of a die, Y = outcome of another die Consider X + 1 What does it mean? We add 1 to all the outcomes of the die (i.e. we now have 2 to 7) ? The probabilities remain unaffected. How does the expected value change? Clearly the mean value will also increase by one. i.e.: πΈ π + 1 = πΈ π +?1 In general: πΈ ππ + π = ππΈ π? + π Modifying Random Variables We often modify the value of random variables. Example: X = outcome of a single throw of a die, Y = outcome of another die Now consider X + Y What does it mean? We consider all possible outcomes of X and Y, and combine them by adding them. The new ? Clearly we need to set of outcomes is 2 to 12. recalculate the probabilities. Uniform Distribution A uniform distribution is one where all outcomes are equally likely. Discrete Example Continuous Example You throw a fair die. What’s the probability of each outcome? 1 π πππ¦ ππ’π‘ππππ = 6 (This ensures the probabilities add up to 1). ? You’re generating a random triangle. You π pick an angle in the range 0 < π < 2 to use to construct your triangle, chosen from a uniform distribution. What is the probability (density) of picking a particular angle? 2 , π π = π π ππ 0 < π < 2 0 ππ‘βπππ€ππ π ? This ensures the area under your PDF π graph (a rectangle with width 2 and 2 height π) is 1. Standard Deviation and Variance Standard Deviation gives a measure of ‘spread’. It can roughly be thought of as the average distance of values from the mean. It’s often represented by the letter π. The variance is the standard deviation squared. i.e. π 2 Variance of a Sample We find the average of the squares of the displacements from the mean. Example: 1cm 4cm 7cm 12cm Mean = 6 Displacements are -5, -2, 1, 6 So variance is: −5 2 + −2 2 + 12 + 62 = 16.5 4 Variance of a Random Variable This is very similar to the sample variation. We’re finding the average of the squared displacement from the mean, i.e.: πππ[π] = πΈ[ π − π 2 ] Using the fact that πΈ ππ + π = ππΈ π + π: πΈ π − π 2 = πΈ π 2 − 2ππ + π2 = πΈ π 2 − πΈ 2ππ + πΈ π2 The expected value of a value is just the value itself. = πΈ π 2 − 2ππΈ π + π2 Since π = = πΈ π 2 − 2πΈ π 2 + πΈ π 2 πΈ[π] = πΈ π2 − πΈ π 2 i.e. We can find the “mean of the squares minus the square of the mean”. Standard Deviation and Variance Example: Find the variance of this biased spinner (which just has the values 1 and 2), represented by the random variable X. k 1 2 P(X=k) 0.6 0.4 πΈ π = 0.6 × 1 + 0.4 × 2 = 1.4 πΈ π 2 = 0.6 × 12 + 0.4 × 22 = 2.2 ? So πππ π = πΈ π 2 − πΈ π 2 = 2.2 − 1.42 = 0.24 STEP Question Fire extinguishers may become faulty at any time after manufacture and are tested annually on the anniversary of manufacture. The time T years after manufacture until a fire extinguisher becomes faulty is modelled by the continuous probability density function: π π‘ = 2π‘ 1 + π‘2 0, 2, πππ π‘ ≥ 0 ππ‘βπππ€ππ π A faulty fire extinguisher will fail an annual test with probability p, in which case it is destroyed immediately. A non-faulty fire extinguisher will always pass the test. All of the annual tests are independent. a) Show that the probability that a randomly chosen fire extinguisher will be destroyed exactly three years after its manufacture is π(5π2 − 13π + 9)/10 (We’ll do part (b) a bit later) b) Find the probability that a randomly chosen fire extinguisher that was destroyed exactly three years after its manufacture was faulty 18 months after its manufacture. What might be going through your head: “I need to consider each of the 3 cases.” “I have a PDF. This requires me to use definite integration.” STEP Question π π‘ = 2π‘ 1 + π‘2 0, 2 , πππ π‘ ≥ 0 ππ‘βπππ€ππ π Since we have a PDF, it makes sense to integrate it so we can find the probability of the extinguisher failing between some range of times. 2π‘ 1 + π‘2 2 ππ‘ = − 1 +π 1 + π‘2 The probability the extinguisher fails sometime in the first year is and during the third year 3 2 − 1 1+π‘ 2 = 1 0 − 1 1+π‘ 2 1 = , during the second year 2 2 1 − 1 1+π‘ 2 = 3 10 1 10 Let’s consider the three cases: a) If it fails during the first year, it must survive the first two tests, before failing the third. This gives a probability of π π − π ππ π b) c) π If it fails during the second year, it must survive the second test and fail on the third, giving π − π π (note ππ that on the first test, the probability of it surviving given it’s not faulty is 1) π If it fails during the third year, then it fails during the third year. We get π. Adding these probabilities together gives us the desired probability. ππ Mean and Variance of Random Variables A point P is chosen (with uniform distribution) on the circle π₯ 2 + π¦ 2 = 1. The random variable π denotes the distance of π from (1,0). Find the mean and variance of X. [Source: STEP1 1987] An important first question is how we could chosen a random point on the circle with uniform distribution. Question: Could we for example choose the x coordinate randomly between -1 and 1, and use π₯ 2 + π¦ 2 = 1 to determine π¦? Click to choose points uniformly across x. No: We can see that because the lines are steeper either side of the circle, we’d likely have less points in these regions, and thus we haven’t chosen a point with uniform distribution around the circle. We’d have a similar problem if we were trying to pick a random point on a sphere, and picked a random latitude/longitude coordinate (we’d favour the poles) Mean and Variance of Random Variables A point P is chosen (with uniform distribution) on the circle π₯ 2 + π¦ 2 = 1. The random variable π denotes the distance of π from (1,0). Find the mean and variance of X. [Source: STEP1 1987] In which case, how can we make sure we pick a point randomly? Introduce a parameter π for the angle anticlockwise from the x-axis say. Clearly this doesn’t give bias to ? certain regions of the arc (satisfying the ‘uniform distribution’ bit). π So what is the distance X? π 2 sin 2 ? (It’s an isosceles triangle, so split into 2) π (1,0) So what is E[X]? 2π π₯π π₯ = 0 π 1 2sin ? 2 2π 1 = π 2π 0 π sin 2 π = π Summary • Random variables have a number of possible outcomes, each with an associated probability. • Random variables can be discrete or continuous. • Discrete random variables have an associated probability mass function. We require that π(π₯) = 1 across the domain of the function (i.e. possible outcomes). • Continuous random variables have an associated probability density function. Unlike ‘conventional’ probabilities, these can have a value greater than 1. We ∞ require that −∞ π π₯ ππ₯ = 1, i.e. the total area under the graph is 1. We can find a probability mass (i.e. the ‘conventional’ kind of probability) by finding the area under the graph in a particular range, using definite integration. • While we have a ‘mean’ for a sample, we have an ‘expected value’ for a random variable, written πΈ[π]. It can be calculated using π₯ π(π₯) for a discrete random variable, and ∞ π₯ π π₯ ππ₯ for a continuous random variable. The expected value for a fair die for −∞ example is 3.5. • The variance gives a measure of spread. For specifically it’s the average squared distance from the mean. We can calculate it using π½ππ πΏ = π¬ πΏπ − π¬ πΏ π , which can be remembered using the mnemonic “mean of the square minus the square of the mean”, or “msmsm”. • πΈ ππ + π = ππΈ π + π. i.e. Scaling our outcomes/adding has the same effect on the mean. Topic 6 – Probability Part 3: Common Distributions Common Distributions We’ve seen so far that can build whatever random variable we like using two essential ingredients: specifying the outcomes, and specifying a PMF/PDF that associates a probability with each outcome. But there’s a number of well-known distributions for which we already have the outcomes and probability function defined: we just need to set some parameters. Bernoulli Multivariate Binomial e.g. Throw of a (possible biased) coin. e.g. Throw of a (possibly biased) die. e.g. Counts the number of heads and tails in 10 throws. Multinomial Poisson Geometric e.g. Counting the number of each face in 10 throws of a die. e.g. Number of cars which pass in the next hour given a known average rate. e.g. The number of times I have to flip a coin before I see a heads. We won’t explore these. Exponential Dirichlet e.g. The possible time before a volcano next erupts. e.g. The possible probability distributions for the throw of a die, given I threw a die 60 times and saw 10 ones, 10 twos, 10 threes, 10 fours, 10 fives and 10 sixes. Bernoulli Distribution The Bernoulli Distribution is perhaps the most simple distribution. It models an experiment with just two outcomes, often referred to as ‘success’ and ‘failure’. It might represent the single throw of a coin. (where ‘Heads’ could represent a ‘success’) Description A single trial with two outcomes. Outcomes “Failure”/”Success”, or {0, 1} ? Parameters? p, the probability ? of success. Probability Function 1−π π₯ =0 π π = π₯ =? π π₯=1 A trial with just two outcomes is known as a Bernoulli Trial. A sequence of Bernoulli Trials (all independent of each other) is known as a Bernoulli Process. An example is repeatedly flipping a coin, and recording the result each time. Binomial Distribution Suppose I flip a biased coin. Let heads be a ‘success’ and tails be a ‘failure’. Let there be a probability π that I have a success in each throw. The Binomial Distribution allows us to determine the probability of a given number of successes in n (Bernoulli) trials, in this case, the number of heads in n throws. Question: If I throw a biased coin (with probability of heads p) 8 times, what is the probability I see 3 heads? H H H T T T T T The probability of this particular sequence is: π3 1 − ?π5 But there’s 83? ways in which we could see 3 heads in 8 throws. Therefore π 3 π»ππππ = 83 π3 1? − π 5 Binomial Distribution Therefore, in general, the probability of k successes in n trials is: π π=π = π π ππ 1 − π π−π Description Outcomes Parameters? Probability Function Binomial D Number of ‘successes’ in n trials. {0, 1, 2, … , n} i.e. between 0 ? and n successes. π, the probability π π π π = π = π 1−π of a single success. π ? of ? π, the number trials We can write B(n,p) to represent the Binomial Distribution, where n and p are the parameters for the number of trials and probability of a single success. If we want some random variable X to use this distribution, we can use πΏ~π©(π, π). The ~ means “has the distribution of”. π−π Frost Real-Life Example While on holiday in Hawaii, I was having lunch with a family, where an unusually high number were left-handed: 5 out of the 8 of us (including myself). I was asked what the probability of this was. (Roughly 10% of the world population is left-handed.) Suppose X is the random variable representing the number of left handed people. Then π~π΅ 8, 0.1 ? 8 × 0.15 × 0.93 5 ? = 1 ππ 2450 πβππππ π π=5 = (This example points out one of the assumptions of the Binomial Distribution: that each trial is independent. But this was unlikely to be the case, since most on the table were related, and left-handedness is in part hereditary. Sometimes when we model a scenario using an ‘off-the-shelf’ distribution, we have to compromise by making simplifying assumptions.) Summary of Distributions so far Similarly, a multivariate distribution represents a single trial with any number of outcomes. A multinomial distribution is a generalisation of the Binomial Distribution, which gives us the probability of counts when we have multiple outcomes. Generalise to n trials Bernoulli e.g. “What’s the probability of getting a Heads?” Binomial e.g. “What’s the probability of getting 3 Heads and 2 Tails?” Generalise to k outcomes Multivariate e.g. “What’s the probability of getting a 5? Generalise to n trials Multinomial e.g. “What’s the probability of rolling 3 sixes, 2 fours and a 1? (Use your combinatorics knowledge to try and work out the probability function for this!) Poisson Distribution Cars pass you on a road at an average rate of 5 cars a minute. What’s the probability that 3 cars will pass you in the next minute? When you have a known average ‘rate’ of events occurring, we can use a Poisson Distribution to model the number of events that occur within that period. We use π to represent the average rate. We can see that when the average rate is 10 (say per minute), we’re most likely to see 10 cars. But technically, we could see a million cars (even if the probability is very low!) k is the number of events (e.g. seeing a car) that occur. Poisson Distribution Assumptions that the Poisson Distribution makes: 1. All events occur independently (e.g. a car passing you doesn’t affect when the next car will pass you). 2. Events occur equally likely at any of time (e.g. we’re not any more likely to see cars at the beginning of the period than at the end) Description Outcomes Number of events occurring within a fixed period given an average rate. {0, 1, 2, … } up to infinity. ? i.e. The Poisson Distribution is a DISCRETE distribution. Parameters? π, the average number of events ? in that period. Probability Function ππ −π π π=π = π ? π! π is Euler’s Number, with the value 2.71… Poisson Distribution Example: An active volcano erupts on average 5 times each year. It’s equally likely to erupt at any time. Q1) What’s the probability that it erupts 10 times next year? 510 −5 π π = 10? = π 10! = 0.018 Q2) What’s the probability that it erupts at all next year? 1−π π =0 50 −5 = 1 −? π 0! = 1 − π −5 = 0.993 Q3) What’s the probability that it next erupts between 2 and 3 years after the current date? i.e. It erupts 0 times in the first year, 0 times in the second year, and at least once the third year. ? −5 −5 −5 π ×π × 1−π = π −10 − π −15 Relationship to the Binomial Distribution Imagine that we segment this fixed period into a number of smaller chunks of time, in each of which an event can occur (which we’ll describe as a ‘success’), or not occur. 1 minute A car passed in this period! A car passed in this period! If we presumed that we only had at most one car passing in each of these smaller periods of time, then we could use a Binomial Distribution to model the total number of cars that pass across 1 minute, because it models the number of successes. Of course, multiple cars could actually pass within each smaller segment of time. How would we fix this? Relationship to the Binomial Distribution We could simply use smaller chunks of time – in the limit, we have tiny slivers of time, so instantaneous that we couldn’t possibly have two cars passing at exactly the same time. 1 minute Now if we’d divided up our time into π chunks where π is large, and we expect an average of π cars to pass, what then is the probability π of a car passing in one chunk of time? (Only Year 8 probability needed!) π π= ? π Therefore, as n becomes infinitely large (so our slivers of time become instantaneous moments), we can use the Binomial Distribution to represent the number of events that occur within some period: π π=π π = lim π→∞ π π π π π 1− π π−π ππ −π = π π! We need some fiddly maths to show this. π tends to arise in maths when we have limits. Uniform Distribution We saw earlier that a uniform distribution is where each outcome is equally likely. Description Each outcome is equally likely. Outcomes x1, x2, …, xn Parameters? None. ? Examples: The throw of a fair die, the throw of a fair coin, the possible lottery numbers this week (presuming the ball machine isn’t biased!). ? Probability Function π π = π₯π = ? 1 π ∀π Geometric Distribution You, Christopher Walken, are captured by the Viet Cong during the Vietnam War, and forced to play Russian Roulette. The gun has 6 slots on the barrel, one of which has a bullet, and the other slots empty. Before each shot, you rotate the barrel randomly, then shoot at your own head. If you survive, you repeat this ordeal. Q1) What’s the probability that you die on the first shot? Q2) What’s the probability that you die on the second shot? π π You survive the first then die?on the second: × π π = π ππ Q3) What’s the probability that you die on the π₯ π‘β shot? p x = π π−π π ? × π π π ? π Geometric Distribution If you have a number of trials, where in each trial you can have a ‘success’ or ‘failure’, and you repeat the trial until you have a success (at which point you stop), then the geometric distribution gives you the probability of succeeding on the 1st trial, the 2nd trial, and so on. Description Succeeding on the xth trial after previously failing. Outcomes Parameters? { 1, 2, 3, … } The trial on which ? you succeed. The probability π of success. ? Probability Function π π₯ = 1−π π₯−1 π ? π Note that if π~πΊπππ(π), then π¬ πΏ = π For example, if we tossed a fair die until we saw a 1, we’d expect to have to throw the die 1 1 ÷ 6 = 6 times on average before we see a 1 (where the count includes the last throw). Side Note: The distribution is called ‘geometric’ because if we were to list out the probabilities for π(1), π(2), π(3) and so on, we’d have a geometric series! Geometric Distribution Tom and Geri have a competition. Initially, each player has one attempt at hitting a target. If one player hits the target and the other does not then the successful player wins. If both players hit the target, or if both players miss the target, then each has another attempt, 4 with the same rules applying. If the probability of Tom hitting the target is always 5 and the 2 probability of Geri hitting the target is always 3, what is the probability that Tom wins the competition? 4 A: ο» 15 4 5 D: ο» B: 8 ο» 15 E: 13 ο» 15 2 3 C: οΌ 4 2 1 1 3 The probability that they both hit or miss is 5 × 3 + 5 × 3 = 5. 4 1 4 So Tom can win by either winning immediately 5 × 3 = 15, or initially 3 4 drawing before winning: 5 × 15, or drawing twice and then winning: 3 2 4 × and so on. This gives us an infinite geometric series with 5 5 4 3 π 2 π = 15 and π = 5. Using 1−π, we get 3. SMC Level 5 Level 4 Level 3 Level 2 Level 1 Frost Real-Life Example My mum (who works at John Lewis), was selling London Olympics ‘trading cards’, of which there were about 200 different cards to collect, and could be bought in packs. Her manager was curious how many cards you would have to buy on average before you collected them all. The problem was passed on to me! (Note: Assume for simplicity that each card is equally likely to be acquired – unlike say ‘Pokemon cards’ [a childhood fad I never got into], where lower numbered cards are rarer) Hint: Perhaps think of the trials needed to collect the next card as a geometric process? Then consider these processes all combined together. Answer: ππππ?cards Explanation on next slide… Frost Real-Life Example Answer: ππππ cards To get the first card, we just need to buy 1 card. 199 To get the second card, each time we buy a card, we have a 200 chance of buying a new card (if not, we keep buying until we have a new one). Since the number of cards we need to buy to get this next card is geometrically distributed, we expected number of cards is 1 200 = . π 199 Combined these expected number of cards we need to buy for each new card, we get 200 200 200 200 + + + β― + 200 199 198 1 π π π = πππ π + π + π + β― + πππ . The bracketed expression is known as a ‘Harmonic Series’, which can be represented as π―πππ . Typing “200 * H(200)” into www.wolframalpha.com got me the answer above. This problem is more generally known as the “Coupon Collector’s Problem” http://en.wikipedia.org/wiki/Coupon_collector%27s_problem Coin Conundrums How would you model a fair coin given you have just a fair die? Solution: Easy! Roll the die. If you get say an even number, declare ? ‘Heads’, else declare ‘Tails’. How would you model a fair die given you have just a fair coin? Solution: A bit harder! Throw the coin 3 times, giving us 8 possible outcomes. Label the first 6 of these outcomes (e.g. HHH, HHT, …). If we get the last two outcomes, then reject these outcomes and repeat. An interesting side question is how many times on average we’d expect to have to throw ? 6 3 the coin. If the probability of being able to stop is p = 8 = 4, the expected value of a 1 4 geometric distribution is π, i.e. 3. Since we throw the coin 3 times each time, then we expect an average of 4 throws. How would you model a fair coin using an unfair coin? Solution: Suppose the probability of Heads on the unfair coin is π. Then throw this coin two times. We have four outcomes: HH, HT, TH and TT, with probabilities π2 , π(1 − π), π(1 − π) and 1 − π 2 respectively. ? Two of these outcomes have the same probability. So declare ‘Heads’ if you threw HT on the biased coin, ‘Tails’ if you threw HT, and repeat otherwise. (You’d expect to have to throw π(π − 1) times on average). Gaussian/Normal Distribution A Gaussian/Normal distribution is a continuous distribution which has a ‘bell-curve’ type shape. It’s useful for modelling variables where the values are clustered, about the mean, and spread out around it with probability dropping off. P(X=x) IQ is a good example. The mean is (by definition) 100, and the probability of having an IQ drops off symmetrically, with Standard Deviation 15 (by definition). 70 85 100 115 130 145 IQ (“Intelligence Quotient”) x Suppose there’s a known mean π and a known standard deviation π. We distribution is denoted as π(π, π 2 ), parameterised by the mean and variance. Then if π~π π, π 2 : π π=π₯ = 1 π 2π π − π₯−π 2 2π2 Z-values We might be interested to know what percentage of the population have an IQ below 130. The z-value is the number of Standard Deviations above the mean. P(X=x) Some rather helpful mathematicians have compiled a table of values that give us π π₯ < π§ , i.e. the probability of being below a particular z-value, for different values of z. This is unsurprisingly known as a z-table. 70 85 100 115 130 145 IQ (“Intelligence Quotient”) x If π = 15 and π = 100, how many Standard Deviations above the mean is 130? Answer = 2 ? Z-values …and our hundredths digit here. We look up the units and tenths digit of our z-value here… So π π₯ ≤ 130 = 0.9772 ? Z-values It’s useful to remember that 68% of values are within 1 s.d. of the mean, 95% within two, and 99.7% within 3 (when the variable is ‘normally distributed’) When scientists referred to a “5π test” needed to officially ‘discover’ the Higgs Boson, they mean that were the data observed to occur ‘by chance’ in the situation where the Higgs Boson didn’t exist (known as the null hypothesis), then the probability is less than that of being 5π away from the mean of a randomly distributed variable: a 1 in 3.5m chance. A Level students studying S2 will encounter ‘hypothesis testing’.