Week 7

Expectation and Linear Transformations

• Let X be a continuous random variable with probability density function (pdf) f(x).
• For any real-valued function g, the expectation is given by
  $E[g(X)] = \int_{-\infty}^{\infty} g(x) f(x) \, dx.$
• Special case: if Y = aX + b, then
  $E[aX + b] = a E[X] + b,$
  since
  $E[aX + b] = \int<em>{-\infty}^{\infty} (a x + b) f(x) \, dx = a \int</em>{-\infty}^{\infty} x f(x) \, dx + b \int_{-\infty}^{\infty} f(x) \, dx = a E[X] + b.$

Nonnegative g(X): an alternative expression for E[g(X)]

• Theorem (nonnegative g): if g(x) ≥ 0 for all x, then
  E[g(X)] = \int_{0}^{\infty} \Pr{g(X) > y} \, dy.
• Sketch of proof:
  • For nonnegative Y with density fY, we have E[Y] = \int{0}^{\infty} \Pr{Y > y} \, dy.
  • Since \Pr{Y > y} = \int{y}^{\infty} fY(x) \, dx,
    interchanging integrals (Tonelli/Fubini) yields
    E[Y] = \int{0}^{\infty} \int{x:g(x) > y} f(x) \, dx \, dy = \int_{-\infty}^{\infty} g(x) f(x) \, dx.
• This justifies using the tail integral representation for nonnegative functions of X.

Example: E[e^X] for X ~ Uniform(0,1)

• Density: $f(x) = \begin{cases}1, &amp; 0 \le x \le 1, \\ 0, &amp; \text{otherwise}.\end{cases}$
• Directly,
  $E[e^X] = \int_{0}^{1} e^{x} \, dx = e - 1.$
• Alternative method via Y = e^X:
  • CDF: for 1 ≤ x ≤ e,
    $F<em>Y(x) = \Pr(Y \le x) = \Pr(e^{X} \le x) = \Pr(X \le \log x) = \int</em>{0}^{\log x} f(t) \, dt = \log x.$
  • Therefore pdf:
    $f<em>Y(x) = (F</em>Y(x))' = \frac{1}{x}, \quad 1 \le x \le e.$
  • Expectation:
    $E[Y] = \int<em>{1}^{e} x \cdot \frac{1}{x} \, dx = \int</em>{1}^{e} 1 \, dx = e - 1.$
• Result: $E[e^{X}] = e - 1$ (same result, different method).

Normal Approximation to Binomial Distributions (DeMoivre–Laplace)

• Let $S<em>n \sim \text{Bin}(n, p).$ For any real numbers a < b, $\lim</em>{n \to \infty} \Pr\left(a \le \frac{S<em>n - np}{\sqrt{np(1-p)}} \le b\right) = \Phi(b) - \Phi(a),$ where $\Phi(z) = \frac{1}{\sqrt{2\pi}} \int</em>{-\infty}^{z} e^{-t^2/2} \; dt.$
• Binomial approximations:
  • Poisson approximation when n is large and p is small: if n p is moderate, use Poisson with parameter $\lambda = np$.
  • Normal approximation is good when $np(1-p)$ is large, typically ≥ 10.

Examples: Binomial approximations and exact calculations

• Example 1: 40 fair coin tosses, X = number of heads. Find $\Pr{X = 20}.$
  • Continuity correction: \Pr{19.5 < X < 20.5}.
  • Standardize: $\frac{19.5 - 20}{\sqrt{40 \cdot 0.5 \cdot 0.5}} = \frac{-0.5}{\sqrt{10}} \approx -0.158,$
    and similarly for 20.5. Thus
    $\Pr{X=20} \approx \Phi(0.158) - \Phi(-0.158) = 2\Phi(0.158) - 1 \approx 0.1272.$
  • Exact value: $\Pr{X=20} = {40 \choose 20} 2^{-40} \approx 0.1254.$
• Example 2: Class size, 450 offers, each attending with probability 0.3. X ~ Bin(450, 0.3). Compute $\Pr{X \ge 151}.$
  • Use continuity correction: $\Pr{X \ge 151} = \Pr{X \ge 150.5}.$
  • With np = 450×0.3 = 135 and Var = 450×0.3×0.7 = 94.5, SD = \sqrt{94.5} \approx 9.725.
  • Z-value: $Z = \frac{150.5 - 135}{\sqrt{94.5}} \approx \frac{15.5}{9.725} \approx 1.59.$
  • Tail probability: $\Pr{Z \ge 1.59} \approx 1 - \Phi(1.59) \approx 0.0559.$
• Example 3: In 10,000 coin tosses, observed heads = 5800. Is the coin fair?
  • If X ~ Bin(10000, 0.5), then E[X] = 5000, Var(X) = 2500, SD = 50.
  • Continuity-corrected: $\Pr{X \ge 5800} = \Pr{X \ge 5799.5}.$
  • Standardize: $Z = \frac{5799.5 - 5000}{50} = \frac{799.5}{50} = 15.99.$
  • Result: probability is effectively 0; thus the coin is not fair in practical terms.

Exponential Distribution

• Density:
  f(x) = \begin{cases} \lambda e^{-\lambda x}, & x \ge 0, \[2pt] 0, & x < 0. \end{cases}
• CDF:
  $F(x) = \Pr(X \le x) = \begin{cases} 1- e^{-\lambda x}, &amp; x \ge 0, \\ 0, &amp; x &lt; 0. \end{cases}$
• Survival function: \Pr(X > x) = 1 - F(x) = e^{-\lambda x}, \quad x \ge 0.
• Parameter: \lambda > 0.
• Moments: If $X \sim \text{Exp}(\lambda)$, then
  $E[X] = \frac{1}{\lambda}, \quad \operatorname{Var}(X) = \frac{1}{\lambda^{2}}.$
• Recurrence for moments: $E[X^{n}] = \frac{n}{\lambda} E[X^{n-1}].$
• Memoryless property: Exponential is memoryless:
  $\Pr{T \ge t+s \,|\, T \ge s} = \Pr{T \ge t} = e^{-\lambda t}.$
• Practical note: Not a good model for human lifetime (memorylessness implies equal remaining life for people at different ages).
• Relation to Poisson process: If events occur as a Poisson process with rate $\lambda$, then the time until the next event T is Exp($\lambda$), with tail \Pr(T > t) = e^{-\lambda t}.

Gamma Distribution

• Motivation: Waiting time until the n-th event in a Poisson process. Let T_n be this waiting time.
• CDF:
  $F<em>{T</em>n}(t) = \Pr(T<em>n \le t) = \Pr( \text{at least } n \text{ events in } [0,t] ) = \sum</em>{k=n}^{\infty} \frac{(\lambda t)^{k}}{k!} e^{-\lambda t} = 1 - \sum_{k=0}^{n-1} \frac{(\lambda t)^{k}}{k!} e^{-\lambda t}.$
• PDF (Gamma density): for shape \alpha > 0 and rate \lambda > 0,
  g(t) = \frac{\lambda^{\alpha}}{\Gamma(\alpha)} t^{\alpha - 1} e^{-\lambda t}, \quad t > 0.
• When $\alpha = 1$, Gamma reduces to the Exponential distribution: $\text{Exp}(\lambda) = \Gamma(1, \lambda).$
• Interpretation: The waiting time until the $\alpha$-th event is Gamma with parameters $(\alpha, \lambda)$; the Gamma variable is the sum of $\alpha$ i.i.d. Exp($\lambda$) variables (only strictly true for integer $\alpha$).
• Applications: modeling arrival times in Poisson processes, time to failures, positive-valued quantities (rainfalls), and financial quantities like insurance claim sizes.
• Gamma function connection: The gamma function is defined as \Gamma(x) = \int_{0}^{\infty} u^{x-1} e^{-u} \, du, \quad x > 0.
  • It satisfies the recursive relation
    $\Gamma(x) = (x-1) \Gamma(x-1).$
  • For positive integers, $\Gamma(n) = (n-1)!.$
• Gamma density with two parameters: shape \alpha > 0 and rate \lambda > 0 is often written as
  g(t) = \frac{\lambda^{\alpha}}{\Gamma(\alpha)} t^{\alpha - 1} e^{-\lambda t}, \quad t > 0.
• Mean and variance of Gamma$(\alpha, \lambda)$:
  • $E[X] = \frac{\alpha}{\lambda}, \quad \operatorname{Var}(X) = \frac{\alpha}{\lambda^{2}}.$
• Important special cases and interpretations:
  • If $X \sim \Gamma(n, \lambda)$ with integer $n$, then $X$ is the sum of $n$ i.i.d. Exp($\lambda$) variables.
  • The Gamma distribution generalizes the exponential distribution; when $\alpha=1$ we recover Exp($\lambda$).
• Example in seismology: times between microearthquakes may be modeled with Gamma vs Exponential; Gamma often provides better fit when inter-event times exhibit variability beyond Poisson assumptions.

Additional notes on Gamma and related concepts

• If $X \sim \Gamma(\alpha, \lambda)$, then the case $\alpha = n$ (a positive integer) corresponds to the sum of $n$ independent Exp($\lambda$) random variables.
• The gamma function appears in the normalization constant of the gamma density; it generalizes the factorial to non-integer values.
• In practice, the gamma distribution is used for positive-valued data with skewness controlled by the shape parameter $\alpha$; larger $\alpha$ yields more symmetric (approximately normal) shapes when scaled appropriately.

Quick recap and connections to foundational principles

• Linearity of expectation: expectations of sums are sums of expectations; transformations of random variables preserve linear relationships under expectation.
• Integral representations and Fubini/Tonelli theorem allow swapping order of integration, enabling tail-integral representations.
• Normal approximation to binomial relies on central limit intuition: binomial counts behave like normal distributions under appropriate scaling.
• Exponential distribution embodies memoryless property, linking to Poisson processes and waiting-time problems.
• Gamma distribution connects to Poisson processes as the waiting time to the n-th event; its properties extend the exponential case to sums of independent exponentials.

Key formulas to remember (LaTeX)

• Expectation of a transformation:
  $E[g(X)] = \int_{-\infty}^{\infty} g(x) f(x) \, dx.$
• Linear transformation of X:
  $E[aX + b] = a E[X] + b.$
• Tail-integral representation (for nonnegative g):
  E[g(X)] = \int_{0}^{\infty} \Pr{g(X) > y} \, dy.
• Normal approximation to Binomial (DeMoivre–Laplace):
  $\lim<em>{n \to \infty} \Pr\left(a \le \frac{S</em>n - np}{\sqrt{np(1-p)}} \le b\right) = \Phi(b) - \Phi(a).$
• Exponential distribution:
  f(x) = \lambda e^{-\lambda x} (x \ge 0), \quad F(x) = 1 - e^{-\lambda x} (x \ge 0), \quad \Pr(X > x) = e^{-\lambda x}.
• Mean and variance of Exp($\lambda$):
  $E[X] = \frac{1}{\lambda}, \quad \operatorname{Var}(X) = \frac{1}{\lambda^{2}}.$
• Memoryless property (Exponential):
  $\Pr{T \ge t+s \mid T \ge s} = \Pr{T \ge t} = e^{-\lambda t}.$
• Gamma density (shape $\alpha$, rate $\lambda$):
  g(t) = \frac{\lambda^{\alpha}}{\Gamma(\alpha)} t^{\alpha - 1} e^{-\lambda t}, \quad t > 0.
• Gamma function:
  \Gamma(x) = \int_{0}^{\infty} u^{x-1} e^{-u} \, du, \quad x > 0,
  $\Gamma(x) = (x-1)\Gamma(x-1), \quad \Gamma(n) = (n-1)! \ (n \in \mathbb{N}).$
• Gamma moments: for $X \sim \Gamma(\alpha, \lambda)$,
  $E[X] = \frac{\alpha}{\lambda}, \quad \operatorname{Var}(X) = \frac{\alpha}{\lambda^{2}}.$