Probability For Machine Learning (Part 1)

Random variables, distributions, and moments

Random variables

• A random variable is a function which assigns a number to events in the sample space. (A better name might be "random-valued function on the sample space.")

$$X = \{1:heads,0:tail\}$$

• We describe the probability of an outcome in terms of the probability of a random variable taking a given value:

$$P(X = 1) = 1/2, P(X^2 = 1) = 1$$

Continuous random variables

• Consider choosing a random number between 0 and 1, where all values are equally likely.
• Since there are an (uncountably) infinite number of values, the probability of any given value is zero:$Prob(X = x_0) = 0$
• Does not mean event is impossible. Points on a line have "measure zero"
• Instead, ask for probability to lie within a given range, e.g.,

$$Prob(a < X<b) = \int_a^b p(x)dx$$

Probability distributions

• More generally, a probability distribution satisfies:
  • $P(x) \geq 0$
  • $\sum_k p(x_k) = 1, x_k$ discrete
  • $\int_{-\infty}^{\infty}p(x)dx=1, x$ continuous

Cumulative distribution function

• The cumulative distribution function, or CDF, gives the probability that X is less than or equal to a given value.

$$F(x) = Prob(X<x) = \int_{-\infty}^x p(x^{'})dx^{'}$$

• It contains (nearly) the same information as the probability density, since

$$p(x) = \frac{dF(x)}{dx}$$

• The CDF is often easier to approximate from empirical data and it is useful since

$$Prob(a < X < b) = F(b) - F(a)$$

Change of variable

• Suppose we wish to change variables, shift the distribution, or consider functions of the random variable. If $x \rightarrow y = y(x)$ then the density in terms of the new variable is given by

$$p(x)dx = g(y)dy$$

which preserves the normalization condition. (This assumes y is an increasing function of x. If not, an absolute value is needed for positivity; and further care is needed if y has critical points.)

$$p(x)dx = g(y)dy$$ $$g(y) = \frac{p(x)}{|dy / dx|}$$

Expectations and moments

• The probability distribution defines weighted averages over the sample space, where the weight of each event is equal to its probability. These are called expected values.

• For the discrete case,

$$E[f(X)] = \sum_{k=1}^n f(x_k)p(x_k)$$

• while for the continuous case,

$$E[f] = \int_{-\infty}^{\infty}f(x)p(x)dx$$

Mean of a distribution

• The mean of the distribution is simply the expectation of the random variable itself:

$$\mu = E[X] = \overline{X} = \langle X \rangle = \left\{ \begin{array}{rcl} \sum_kx_kp(x_k) \\ \int xp(x)dx \end{array}\right.$$

• In the case of an infinite sample space, whether continuous or discrete, the mean is not guaranteed to exist since the integral or the sum might not converge.

Moments of a distribution

• The moments of a distribution are the expectation of powers of the random variable itself.

$$\mu_l \equiv E[X^l] \equiv \langle X^l \rangle = \left\{ \begin{array}{rcl} \sum_kx_k^lp(x_k) \\ \int x^lp(x)dx \end{array}\right.$$

• If all the moments are known – and if they exist – they can be used to get the expectation of other functions using the linearity of the expectation operator

$$E[cf(X)] = cE[f(x)]$$ $$E[f(x) + g(X)] = E[f(x)] + E[g(x)]$$

Variance and standard deviation

• Of particular interest is the second moment, in combination with the mean, defining the variance:

$$\sigma^2 = Var(X) = E[(X-\mu)^2] = E[X^2] - E[X]^2$$

• The standard deviation, which is the square root of the variance, has the same units as the random variable (e.g., rate of return, dollars, etc.)

Higher moments characterize properties of a distribution

• Variance – dispersion measure based on second moment

$$\sigma^2 \equiv E[(X-\mu)^2] = \int (x-\mu)^2p(x)dx$$

• Skewness – asymmetry parameter based on 3rd moments; dimensionless – normalized cumulant

$$s \equiv \frac{E[(X-\mu)^3]}{\sigma^3} = E[(\frac{X-\mu}{\sigma})^3]$$

• Kurtosis – measure of tail "weights" in terms of 4th moments; zero for Gaussian, bounded below by -1.

$$\kappa \equiv \frac{E[(X-\mu)^4]}{\sigma^4} - 3$$

Covariance and correlation

• For any two random variables, not necessarily independent or identically distributed, their covariance is defined as

$$Cov(X,Y) \equiv E[(X - \mu_x)(Y - \mu_y)] = E[XY] - \mu_x \mu_y$$

• The correlation is proportional to the covariance,

$$\rho(X,Y) = Corr(X,Y) = \frac{Cov(X,Y)}{\sqrt{Var(X)Var(Y)}} = E[(\frac{X-\mu_x}{\sigma_x})(\frac{Y - \mu_y}{\sigma_y})]$$

• Dividing the covariance by the standard deviations makes the correlation a pure number, and

$$-1 \leq \rho(X,Y) \leq +1$$

Summary

• "Random variables" are functions that assign a number to events in the sample space. They can be discrete, continuous, or a mix of both.
• The probability distribution is positive and sums to one.
• Expected values, or expectations, are weighted averages that use the probabilities as the weights.
• The moments of a distribution, such as the mean, are expectations of various powers of a random variable. They are numbers, not functions.
• Variance, skewness, and kurtosis are simple functions of the moments that characterize the shape of the probability distribution.
• When there are multiple random variables, their covariance and correlation are also computed as expectations.

(To be Continued!)