Continuous Distributions (Uniform, Normal, Exponential) Powerpoint

Home , Exponential distribution, Poisson distribution

• An integrable function f : R → [0,1] such that ∫Rf(x)dx = 1 is called a probability density function (pdf) •The distribution function for the pdf is given by x F(x) = ∫ f(z)dz -∞ and corresponds to the cumulative distribution function for the discrete case • Sampling from the distribution corresponds to solving the equation rsample x = ∫ f(z)dz -∞ for rsample given random probability values 0 ≤ x ≤ 1 © 2008 Winton 3

Uniform Distribution

probability density function (area under the curve = 1) p(x)

1 (b-a)

rsample

•The pdf for values uniformly distributed across [a,b] is given by 1 f(x) = (b-a) © 2008 Winton 4 Sampling from the Uniform Distribution • (pseudo)random numbers x drawn from [0,1] distribute uniformly across the unit interval, so it is evident that the corresponding values rsample = a + x(b-a) rsample slope = (b-a) will distribute uniformly across [a,b] • Directly solving b

rsample x = f(z)dz ∫ a -∞ x for rsample as per 0 1

•The mean μ of the uniform distribution is given by

b ⎛ 1 ⎞ b2 − a 2 1 b + a μ = E(X) = ∫ z ⎜ ⎟ dz = = (midpoint of [a, b] ) a ⎝ b - a ⎠ 2 b - a 2 z f(z) dz

•The standard deviation σ of the uniform distribution is obtained from the variance σ2 where 2 b ⎛ b + a ⎞ ⎛ 1 ⎞ (b - a)2 σ2 = E((X -μ)2 ) = ∫ ⎜z - ⎟ ⎜ ⎟ dz = (with some work) a ⎝ 2 ⎠ ⎝ b - a ⎠ 12 (z-μ)2 f(z) dz © 2008 Winton 6 Normal Distribution

• For a finite population the mean (m) and standard deviation (s) provide a measure of average value and degree of variation from the average value • If random samples of size n are drawn from the population, then it can be shown (the Central Limit Theorem) that the distribution of the sample means approximates that of a distribution with mean: μ = m and σ. s standard deviation: σ = n pdf: (x−μ)2 − 1 2 f(x) = e 2σ σ 2π which is called the Normal Distribution • The pdf is characterized by its "bell-shaped" curve, typical of phenomena that distribute symmetrically around the mean value in decreasing quantity as one moves away from the mean © 2008 Winton 7 Empirical Rule for Normal Distribution • The "empirical rule" is that – approximately 68% of sample values are in the interval [μ-σ,μ+σ] – approximately 95% are in the interval [μ-2σ,μ+2σ] – almost all are in the interval [μ-3σ,μ+3σ] • This says that if n is large enough, then a sample mean for the population is accurate with a high degree of confidence, since σ decreases with n – What constitutes "large enough" is largely a function of the underlying population distribution – The Central Limit Theorem assumes that the samples of size n which are used to produce sample means are drawn in a random fashion © 2008 Winton 8 Measurements Found to Distribute Normally

• Hogg and Craig (Introduction to Mathematical Statistics) cites as distributing normally such disparate phenomena as – The diameter of the hole made by a drill press – The score on a test – The yield of grain for a plot of ground – The length of a newborn child • The assumption that grades on a test distribute normally is the basis for so-called "curving" of grades – This assumes some underlying random phenomena controls the measure given by a test • e.g., genetic selection • The practice sometimes is used to assign grades of A,B,C,D,F based on how many "standard deviations" separates a percentile score from the mean – If the mean score is 77.5, and the standard deviation is 8, then the curve of the class scores would be given by • A: 94 and up (2.5%) • B: 86-93 (13.5%) • C: 70-85 (68%) • D: 62-69 (13.5%) •F: otherwise(2.5%) – Most people "pass", but A's are hard to get • This could be pretty distressing if the mean is 95 and the standard deviation is 2 (i.e., 90 is an F) © 2008 Winton 9

Normal Distribution: Example Plot

• Plot with with μ = 30 and σ = 10 – The distribution is completely determined by knowing the value of μ and σ

f(x)

• The standard normal distribution is given by μ = 0 and σ = 1, in which case the pdf becomes x2 1 − f(x) = e 2 2π

nsample

• The linear relationship nsample rsample = μ + σ⋅nsample between a normal distribution and the standard normal distribution means it is sufficient to sample from the standard normal distribution • There is no "closed-form formula" for nsample, so approximation 0 x techniques for inverting the .51 cumulative distribution function or a method such as accept-reject has to be used substitute z=µ+σ·t dz = σdt

• The exponential distribution arises in connection with Poisson processes – A Poisson process is one exhibiting a random arrival pattern in the following sense: • For a small time interval Δt, the probability of an arrival during Δt is λΔt, where λ = the mean arrival rate • The probability of more than one arrival during Δt is negligible • Interarrival times are independent of each other. – This is a kind of "stochastic" process, one for which events occur in a random fashion • Under these assumptions, it can be shown that the pdf for the distribution of interarrival times is given by f(x) = λe-λ x which is the exponential distribution © 2008 Winton 13 Properties of the Exponential Distribution

• If it can be shown that the number of arrivals during an interval is Poisson distributed (i.e., the arrival times are Poisson distributed), then the interarrival times are exponentially distributed – The mean arrival rate is given by λ and the mean interarrival time is given by 1/λ – The Poisson distribution is a discrete distribution closely related to the binomial distribution and will be considered later • It can be shown for the exponential distribution that the mean is equal to the standard deviation; i.e., – μ = σ = 1/λ • The exponential distribution is the only continuous distribution that is "memoryless", in the sense that P(X > a+b | X > a) = P(X > b) © 2008 Winton 14 Graphical Appearance of the Exponential Distribution f(x)

f(x)=λe-λ x

• The case λ = 1 gives the standard exponential distribution – Inverting is straight forward since nsample nsample e−zdz = - e-z = 1 - e-nsample = x so nsample = -log (1-x) ∫ 0 e 0 – closed form formula for obtaining a normalized sample value (nsample) using a random probability x • General sample values (rsample) are obtained from the standard exponential distribution by 1 1 rsample = nsample = - log (1- x) λ λ e © 2008 Winton 16 Sampling Function for the Standard Exponential Distribution

rsample

1 - log (1 - x) λ e

1/λ

x 0 (e-1)/e 1 © 2008 Winton 17 Considerations in Using the Exponential Distribution • The utility of the exponential distribution for discrete systems simulation is its effectiveness for modeling a random arrival pattern – Implicit characteristic of a Poisson process – Sampling for interarrival times is a natural approach for introducing new items into the model one at a time – Care must be taken that when used for this purpose, the exponential distribution is applied to relatively short time periods during which the arrival rate is not dependent on time of day • For example, the model could progress in 1 hour service intervals representing slow, moderate, and peak demand, each governed by an exponential distribution with an appropriate mean interarrival time • A more sophisticated approach is to adjust the arrival rates dynamically with time, a concept studied under the topic of joint probability distributions © 2008 Winton 18 Verfication that μ = σ for the Exponential Distribution ∞ ∞ ∞ -1 ∞ 1 μ = E(X) = ∫ z λe-λz dz = - ze-λz + ∫ e-λzdz = 0 + e−λx = 0 0 0 λ 0 λ u dv (integration by parts) uv - ∫ vdu ∞ ∞ ∞ σ2 = E()()X - 1/λ 2 = ∫()z - 1/λ 2 λe-λzdz = - z - 1/λ ()2 e-λz - ∫ - e-λz 2()z - 1/λ dz 0 0 0 u dv 1 2 ∞ 2 ∞ 1 = + z λ e−λzdz - λe−λzdz = 2 ∫ 2 ∫ 2 λ λ 0 λ 0 λ = 1/λ = 1 (by definition) • For the pdf of the exponential distribution note that f(x) = λe-λ x f’(x) = - λ2 e-λ x so f(0) = λ and f’(0) = - λ 2 • Hence, if λ < 1 the curve starts lower and flatter than for the standard exponential. The asymptotic limit is the x-axis.