Sample Spaces, Random Variables 2012 1 Probabilities In talking about probabilities, the fundamental object is Ω, the sample space. Points (elements) in Ω are denoted (generically) by ω. We assign probabilities to subsets of Ω. Assume for the moment that Ω is finite or countably infinite. Thus Ω could be the space of all possible outcomes when a coin is tossed three times in a row or say, the set of positive integers. A probability P is then a function from the power set (the class of all possible subsets) of Ω, which we will denote by A, to the interval [0, 1] satisfying the following properties: • (i) P (Ω) = 1 . • (ii) P (φ) = 0 . • (ii) If {An } is a sequence of mutually disjoint subsets of Ω, then, P (∪∞ i=1 Ai ) = ∞ X P (Ai ) . i=1 In statistical/practical contexts it helps to think of Ω as the set of all possible outcomes of a chance (random) experiment. Probabilities assigned to subsets of Ω are governed by the nature of the chance mechanism (like tossing a coin, throwing a die, shuffling a pack of cards etc.) 1 Example: Consider a fair coin tossed n times, where we record H for the ith trial if the coin turns up Heads and T otherwise. The sample space Ω for this random experiment is the set of all sequences of H’s and T ’s that have length n. The size of Ωis 2n . As far as the assignment of probabilities is concerned, let ω denote a generic sequence of Hs and T s. If the outcomes of the various trials are independent of each other P ({ω}) = 1/2n . Consider now the event E that the 1st toss yields an H. Since we are tossing a fair coin and the trials are independent, the probability of this is readily seen to be 1/2. Let’s now compute the probability a different way. Now E = ∪ω begins with H {ω} and, therefore, X P (E) = ω P ({ω}) = begins with 1 2n−1 = . n 2 2 H In general however Ω can be uncountably infinite (for example when Ω is [0, 1]) in which case (for certain technical reasons that we do not need to go into) A is not taken to be the entire power set, but is chosen to be a smaller class. Thus we do not assign probabilities to all subsets of Ω, but only to those that belong to A. The class A is assumed to contain Ω and to be closed under complementation and countable union. P is then, a function from A to [0, 1] with the same properties as above. The properties (i), (ii) and (iii) are called the basic axioms of probability. From these properties, it is easy to deduce the following: • (a) For finitely many disjoint sets A1 , A2 , . . . , An , P (∪ni=1 Ai ) = n X P (Ai ) . i=1 • (b) P (A) = 1 − P (Ac ) . • (c) If A ⊂ B, then P (A) ≤ P (B). • (d) For any two sets A and B (not necessarily disjoint), P (A ∪ B) = P (A) + P (B) − P (A ∩ B) . Exercise: Using (i), (ii), (iii), deduce (a), (b), (c) and (d). There is an interesting generalization of (d) to the case of more than 2 sets. 2 Proposition: For (not necessarily disjoint) sets A1 , A2 , . . . An , X X X P (∪ Ai ) = P (Ai ) − P (Ai ∩ Aj ) + P (Ai ∩ Aj ∩ Ak ) i 0. The conditional probability of A | B (read, A given B) is, P (A | B) = P (A ∩ B) . P (B) It is not difficult to show that this defines a valid probability. In other words if we define a function PB from A to [0, 1] as follows: PB (A) = P (A | B) , then PB satisfies the basic axioms of probability. The key idea behind defining conditional probabilities is to update the probabilities of events given the knowledge that B happened. Note that for any event C ⊂ B c , P (C | B) = 0, which is in accordance with common sense. On the other hand, for any C ⊂ B, P (C | B) ≥ P (C) ; in other words, the conditional probability of a sub-event of B, given that B happened is generally higher than the unconditional probability. Joint probabilities can be expressed in terms of conditional and marginal probabilities in the following manner: P (A ∩ B) = P (A | B) P (B) , provided P (B) > 0. Also, P (A ∩ B) = P (B | A) P (A) , provided P (A) > 0. More generally, for sets A1 , A2 , . . . An , we have, P (∩ni=1 Ai ) = P (A1 ) P (A2 | A1 ) P (A3 | A1 ∩ A2 ) . . . P (An | ∩n−1 i=1 Ai ) . One of the most important consequences of the notion of conditional probability is Bayes Rule. Simple as it looks, its ramifications are profound. Bayes Rule gives us a way of obtaining the conditional probability of one event given another, when we know the marginal probabilities and the conditional probability of the second event given the first. We state Bayes Rule in the form of the following theorem. Theorem: Suppose that {Bj } is a (finite/infinite) partition of Ω and A is any event 4 with P (A) > 0. Also suppose that P (Bi ) > 0 for each i (without any loss of generality). Then, P (A | Bj ) P (Bj ) P (Bj ∩ A) =P . P (Bj | A) = P (A) P (A | Bi ) P (Bi ) Illustrating Bayes Rule: There are three cabinets, A, B and C, each of which has two drawers. Each drawer has one coin. A has two gold coins, B has two silver coins and C has one gold and one silver coin. A cabinet is chosen at random; one drawer is opened and a silver coin found. What is the chance that the other drawer also has a silver coin ? To fit this in the framework of the above theorem, take the Bj ’s to be the events that A is chosen, B is chosen and C is chosen. For brevity we shall denote the events by A, B and C. Note that these events are indeed a partition of the sure event that a cabinet is chosen, and furthermore P (A) = P (B) = P (C) = 1/3 . Let S1 denote the event that the opened drawer of the chosen cabinet has a silver coin. Clearly, we are required to find P (B | S1). Now, using Bayes Rule, we have P (S1 | B) P (B) P (S1 | A) P (A) + P (S1 | B) P (B) + P (S1 | C) P (C) 1 × 1/3 = 0 × 1/3 + 1 × 1/3 + 1/2 × 1/3 1 = 1 + 1/2 = 2/3 . P (B | S1) = Independence of events: Events A and B are said to be independent if P (A ∩ B) = P (A) × P (B). In general, events A1 , A2 , . . . , An are said to be mutually independent if for any subcollection {Ai1 , Ai2 , . . . , Aik } of {A1 , A2 , . . . , An }, it is the case that, P (Ai1 ∩ Ai2 ∩ . . . Aik ) = P (Ai1 ).P (Ai2 ). · · · P (Aik ) . Exercise: If A1 , A2 , . . . , An are independent, then so are B1 , B2 , . . . Bn where each Bi is either Ai or Aci . Consider events A and B with P (A) > 0, P (B) > 0. and B is equivalent to saying that, P (A | B) = P (A) , and equivalently, P (B | A) = P (B) . 5 Then the independence of A In other words knowledge about B does not affect the probability of the event A and viceversa, which is intuitively compatible with the notion of (stochastic) independence. 2 Random Variables A random variable X is a function from Ω to the real numbers. Thus, for each ω, X(ω) is a real number. In talking about the value of a random variable at a particular sample point, the argument ω is usually suppressed. Most of probability and statistics deals with the study of random variables. Random variables can broadly be classified into two types. These are (i) Discrete and (ii) Continuous. Discrete random variables are those that assume finitely many or countably many values with positive probability. On the other hand, for a continuous random variable X the probability that X assumes a particular value x is 0 for any x. However, the probability that X lives in some interval [a, b] can be positive; in fact, the probability that X lives in R is 1. This apparently smells paradoxical; if the chance of any particular value in [a, b] being realized is 0, how can the probability of an interval [a, b] even be positive? This is where the subtlety of the theory lies. An interval contains uncountably many points and the addition rule for probabilities only holds for finitely or countably many events; thus the probability of an interval cannot be computed by adding up the probability of each point in that interval, namely 0. Yet chances can be meaningfully assigned to subsets of R and be meaningfully interpreted. If we couldn’t do this, there wouldn’t have been much use for the theory. Example 1: This is an example of a discrete random variable. Consider the coin-tossing experiment, where a coin is flipped once. The sample space Ω = {0, 1}. Let X be the random variable that assumes the value 1 if heads comes up, and 0 if tails comes up. Thus X(H) = 1 and X(T ) = 0. If the chance of the coin landing heads up is p, then clearly, P (X = 1) = p and P (X = 0) = q = 1 − p . The random variable X is called a Bernoulli random variable. Example 2: This is also a discrete random variable that is composed by adding independent Bernoulli random variables. Suppose you flip a coin n times and record the outcomes of each toss. The space of all possibilities is the sample space Ω and is made up of sequences of length n where each slot is either H (heads) or T (tails). What is the size of Ω? The random variable X records the number of heads in a sequence. Thus if ω denotes a sequence, then X(ω) = number of heads in ω. Thus X can assume any value between 0 6 and n. It is easy to see that, P ({ω}) = pX(ω) q n−X(ω) , and consequently, n P (X = m) = pm q n−m . m Note that X is the sum of n independent Bernoulli random variables. Example 3: This is a discrete random variable that takes infinitely many values unlike the previous two. Suppose, we keep on flipping a coin, till we get a heads and then stop. Let the probability of the coin landing heads up in a flip be p. The sample space is then given by Ω = {H, T H, T T H, T T T H, · · ··} . Let X denote the number of flips needed to get the first head. Clearly X can be any positive integer. Then X is clearly a random variable and it is easily seen that, P (X = m) = q m−1 p . The geometric distribution has an interesting property called the memoryless property. This can be stated as follows: Let X be a geometric random variable. Given any two positive integers m and n, P (X > m + n | X > m) = P (X > n) . The reason why the above is referred to as the memoryless property is fairly clear. Given the information that you haven’t seen a head in the first m tosses, the chance that you won’t see a head in n further tosses, is the same as the probability that you don’t see a head in the first n tosses. So in a sense you are starting from square one, from the m + 1’st toss onwards. A distribution that arises very naturally from the geometric is the negative binomial distribution. Often useful in analyzing discrete data, it can be described very simply, in terms of the following random experiment: Suppose I keep tossing a coin till I get the r’th heads (success). Let X be the number of tails that I get before the r’th heads. Clearly X is a random variable. What is the chance that X = x? To compute this, note that for X to be equal to x, there must have been a total of x + r tosses including the r’th heads and in the first x + r − 1 tosses there must have been r − 1 heads. Now, the total number of (distinct) sequences of H’s and T ’s that are r long x +x+r−1 and end in an H and have r − 1 heads in the first x + r − 1 slots is just x+r−1 ≡ . The chance than any such sequence is actually x r−1 r x realized is just p q . Hence, x+r−1 r x P (X = x) = p q . x 7 Note that X can assume any value greater than or equal to 0 with positive probability. Exercise: If W = X + r is the total number of tosses needed to get r heads, write down the p.m.f. of W and show that X1 + X2 + . . . + Xr where the Xi ’s are i.i.d. geometric (as defined above) is distributed like W . Exercise: Prove that the geometric distribution satisfies the memoryless property. The interesting fact is that among all discrete random variables supported on {1, 2, 3, . . .} the geometric distribution is the only one that satisfies the memoryless property. In other word, the memoryless property completely characterizes the geometric distribution. Other examples of discrete random variables include Poisson, Negative Binomial, etc. The probability mass function of a discrete random variable x taking values in, say, {0, 1, 2, 3, . . .} is the function p from the set {0, 1, 2, 3, . . .} to the set [0, 1] defined by, p(i) = P (X = i) . Example 4: This is an example of a continuous random variable. Let the sample space Ω = (0, 1) and let P be a probability defined (uniquely) on Ω in the following way. For any 0 < x < 1, P ((0, x]) = x . This is a good probability model for drawing a point at random from the interval (0, 1). What it says is that the chance of a randomly chosen point belonging to an interval is precisely given by the length of the interval. Let U be a random variable defined as follows: U (ω) = ω . Then, it is easy to see that, P (U ≤ u) = u , for any 0 < u < 1. The random variable U is said to have the uniform distribution on (0, 1). In some sense, U is the most basic or fundamental random variable, because, as we shall see presently, all random variables can be generated from U . The Distribution Function: The probabilistic behavior of a random variable is captured completely by its distribution function. If X is a random variable, it’s distribution function FX , henceforth denoted by F , is defined by, F (x) = P (X ≤ x) . F has the following properties. 8 • (i) 0 ≤ F (x) ≤ 1 . • (ii) If x < y, then F (x) ≤ F (y). Also, lim F (y) = F (x) . y→x+ In other words F is monotone increasing and right continuous. • (iii) limx→−∞ F (x) = 0 and limx→∞ F (x) = 1 . F need not be left-continuous as can be shown by counter-example. Formally, a continuous random variable is defined as one whose distribution function is continuous. Quantiles and Inverse Distribution Functions: For any 0 < p < 1, the p’th quantile of F is any number xp such that F (xp −) ≡ lim F (y) = P (X < xp ) ≤ p ≤ P (X ≤ xp ) = F (xp ) . y→xp − Clearly, the p’th quantile need not be unique. However, if F is a strictly increasing continuous function, in which case it is one-one, it’s inverse function F −1 is well defined on (0, 1). For any 0 < p < 1, F −1 (p) is the unique number x, such that F (x) = p, and F −1 (p) is the unique p’th quantile of F . When p = 0.5, we refer to the quantile as the median. Fortunately, there is a neat way to define the inverse function F −1 even when F is not strictly increasing and continuous. For any 0 < t < 1, we set, F −1 (t) = smallest element of {x : F (x) ≥ t} . The fact that the above set does indeed have a smallest element can be shown. It is now easy to see that F −1 (t) is indeed a t’th quantile of F , though not necessarily the unique one. Note that, F −1 (t) ≤ x ⇔ F (x) ≥ t . We can now state a crucial theorem. Theorem: Let X be a random variable with distribution function F . Let U be a Uniform random variable on (0, 1). Then Y = F −1 (U ) is also a random variable and its distribution function is F . Proof: We have, P (F −1 (U ) ≤ x) = P (F (x) ≥ U ) = P (U ≤ F (x)) = F (x) . Thus, by knowing F , and hence in principle F −1 , one can generate a random variable with distribution function F , provided one can generate from a U (0, 1) distribution. Another 9 related theorem follows. Theorem: If X is a continuous random variable, then F (X) has the uniform distribution. The proof of this theorem, in the case that F is strictly increasing and continuous, is given in Rice. The general case is omitted. Probability mass functions and densities: For discrete random variables, we can obtain the distribution function F from the mass function p in the following way: Suppose that X is a discrete random variable taking on the values x1 , x2 , x3 , . . . with probabilities p1 , p2 , p3 , . . . , . If p denotes the probability mass function, then p(xi ) = pi . The distribution function F of X is then given by, X F (x) = p(xi ) . i:xi ≤x Many continuous distribution functions (and in fact, the vast majority of those which we deal with in statistics) possess a probability density function, which is basically the analogue of the probability mass function in the discrete case. A continuous distribution function F is said to possess a density f if there exists a non-negative real-valued function f , such that for all a < b, Z b F (b) − F (a) = P (a < X ≤ b) = f (x) dx . a If f is continuous, then it is easy to see that f (x) is the derivative of F at x. Also, note that the integral of f over the entire line must be 1; i,e. Z ∞ f (x) dx = 1 . −∞ The uniform distribution on (0, 1) is easily seen to have density function f , given by f (x) = 1 whenever 0 < x < 1 and 0 elsewhere. Note that if f is a density function for a distribution function F , and g is obtained by changing the values of f at finitely many points, then g is also a density function. Many distributions are specified using density functions rather than the distribution functions themselves. In particular, one of the most crucial distributions in statistics, the normal distribution, which arises all over the place, is specified through it’s density. The distribution function of the normal cannot be written in closed form. Recall that a normal 10 random variable with mean µ and standard deviation σ has density, 1 1 2 φµ,σ (x) = √ exp − 2 (x − µ) . 2σ 2πσ The exponential distribution, which is used to model waiting times, has density function given by, f (x) = λ exp(−λ x) , where λ is some positive constant. Its distribution function is available in closed form as, F (x) = 1 − exp(−λ x) . Among all continuous distributions on (0, ∞), the exponential distribution is the only one that has the memoryless property. Exercise: Show that the exponential distribution has the memoryless property; i.e. if X is an exponential random variable and s, t > 0, then, P (X > s + t | X > s) = P (X > t) . The converse is more difficult to prove, but a fun exercise. We’ll skip it for now. Exercise: Given a uniform random variable U , how would you generate from an exponential distribution with parameter λ ? Transforming Random Variables: In statistics, we are often interested in deriving the distribution of a given function of a random variable. Thus, if X is a random variable, we might be interested in finding the distribution of g(X), where g is a fixed pre-specified function. We consider, first, the case when X is discrete. Theorem: Suppose X is a discrete random variable assuming values {x1 , x2 , x3 , . . .} with probabilities {p1 , p2 , p3 , . . .}. Let Y = g(X), where g is a given function and let {y1 , y2 , y3 , . . .} be the values assumed by Y . Then the p.m.f. of Y is given by: X P (Y = yi ) = pj . j:g(xj )=yi The proof is immediate. For continuous random variables with a density f , the Jacobian Theorem gives us 11 the distribution of a transformed variable under some regularity conditions on the transformation g. We state the theorem below (without proof). Change of variable theorem: Let X be a continuous random variable with density function f and g be a real-valued function defined on some open interval I, such that 0 P (X ∈ I) = 1. Assume further that g is continuously differentiable and that g (x) 6= 0 for x ∈ I(these assumptions actually entail that g is a strictly monotone transformation on I). Let Y = g(X). Then the density function of Y can be computed as, fY (y) = f (g −1 (y)) | 1 d −1 g (y) |= f (g −1 (y)) | 0 −1 |, y ∈ g(I) , dy g (g (y) and 0 otherwise. Proof: Suppose I = (a, b) and g is increasing. Then a < X < b with probability 1 and the density f can be taken to be identically 0, outside of I. Also g(a) < Y ≡ g(X) < g(b) with probability 1 and the density fY of Y can be taken to be identically 0 outside of g(I). Let FY denote the distribution function of Y . Let g(a) < y < g(b). Then, FY (y) = P (Y ≤ y) = P (g(a) ≤ Y ≡ g(X) ≤ y) = P (a ≤ g −1 (Y ) ≤ g −1 (y)) . Since, X ≡ g −1 (Y ), FY (y) = P (a ≤ X ≤ g −1 (y)) = P (X ≤ g −1 (y)) = FX (g −1 (y)) . On differentiating the above, with respect to y, we obtain, fY (y) = f (g −1 (y)) d −1 1 g (y) = f (g −1 (y)) 0 −1 , dy g (g (y) where the last equality follows from the chain rule and the fact that g and g −1 have positive non-vanishing derivatives. 2 The equality in the above display clearly implies that for any subset S of I: Z Z 1 | dy . f (x) dx = f (g −1 (y)) | 0 −1 g (g (y) S g(S) (2.1) In fact, the above formula is valid very generally; f does not need to be a density function. Any non-negative function, or for that matter, any function, with a well-defined (finite) integral will do. Let’s see how we can apply this formula to compute an integral, that one might come across in calculus. Suppose we wish to compute, Z 1 S= u sin u2 du . 0 12 We write Z 1 Z 2 1 u sin u du = 0 0 1 2u sin(u2 ) du , 2 2 and then set w = u , noting that as u runs from 0 t0 1, so does w in the same direction. But w = u2 means that dw = 2u du and thus Z 1 Z 1 1 1 2 2u sin(u ) du = sin w dw , 0 2 0 2 by direct substitution and we obtain, S= 1 (1 − cos 1) . 2 Basically, what we have done here is use (??) informally. To see this, let f (x) = 1 2x sin x2 I = (0, 1) and g(x) = x2 . 2 0 0 Clearly, g is continuously differentiable on I and g is non-vanishing (since g (x) = 2x)). Now, the inverse transformation g −1 from g(I) = (0, 1) to I is g −1 (y) = Also, f (g −1 (y)) = √ y. 1 √ 2 w sin w . 2 Thus, Z S = f (x) dx ZI = g(I) Z 1 = 0 f (g −1 (y)) | 1 g 0 (g −1 (y) 1 1 √ 2 y sin y √ 2 2 y Z 1 1 = sin y dy 2 0 1 = (1 − cos 1) . 2 13 | dy

Sample Spaces, Random Variables

Related Documents

Recommended Documents

Description

Related Documents

Free up your schedule!

Take 5 seconds to unlock