Ex. 1 – Show that \(x^2 + exp(x) + 2x^4 + 1\) is convex.
This function f(x) is convex if the following is true:
\[f(\alpha x + \beta y) \leq \alpha f(x) + \beta f(y), \alpha \geq 0, \beta \geq 0, \alpha + \beta = 1\]
However, the consequent equation is time-consuming to simplify. One property of convex functions makes determining this easier:
If \(f_1\) and \(f_2\) are convex, then \(\alpha f_1 + \beta f_2\) is also convex for any \(\alpha, \beta \geq 0\).
Consider $f_1 = x^2, \(f_2 = 2x^4 + 1\), and \(f_3 = exp(x)\):
\[f_1: (\alpha x + \beta y)^2 \leq \alpha x^2 + \beta y^2\] A similar quadratic funciton has larady been solved in the lesson, but to recap:
\[ 0 \leq \alpha x^2 + \beta y^2 - (\alpha x + \beta y)^2 \]
\[ 0 \leq \alpha x^2 - \alpha^2 x^2 + \beta y^2 - \beta ^2 y^2 - 2\alpha \beta xy\] Factor out \(\alpha\) and \(\beta\) terms: \[ 0 \leq \alpha x^2 ( 1 - \alpha ) + \beta y^2 (1 - \beta) - 2\alpha \beta xy = \alpha \beta x^2 + \alpha \beta y^2 - 2\alpha \beta xy \]
\[ \alpha \beta (x^2 - 2xy - y^2) = \alpha \beta (x - y)^2 \geq 0\]
This inequality is always true for real numbers. Next, consider whether \(exp(x)\) is convex:
\[f_2: exp(\alpha x + \beta y) \leq \alpha exp(x) + \beta exp(y)\]
\[ 0 \leq \alpha exp(x) + \beta exp(y) - exp(\alpha x + \beta y)\]
Divide by \(exp(x)\), which is permissible as this can never be equal to zero:
\[ 0 \leq \alpha + \beta exp(y - x) - exp(\alpha x + \beta y - x) = \alpha + \beta exp(y - x) - exp(\beta y - x(1-\alpha))\]
\[ 0 \leq = \alpha + \beta exp(y - x) - exp(\beta (y - x)) = \alpha -exp(\frac{(y-x)^\beta}{\beta(y-x)}) = \alpha + \alpha exp(\frac{(y-x)}{\beta})\]
After much simplifying, we’re left with two terms, both of which are non-negative for all real numbers. Therefore exp(x) is convex.
Finally, \(f_3 = x^4 + 1\) is convex owing to the property of convex functions as follows:
If f(x) and g(x) are convex, then f(g(x))s convex under non-decreasing conditions. The f(x) and g(x) in this question are \(f(x) = x^2\) and \(g(x) = x^2 +1\).
Taken all together, the original function is convex.
Ex. 2 – Show that the mean of the exponential distribution
\[p(x) = \begin{cases} \ \lambda e^{-\lambda x}, & 0 < \lambda \leq x \\ 0, & x < 0 \end{cases}\]
is \(\mu = \frac{1}{\lambda}\) and its variance is \(\sigma ^2 = \frac{1}{\lambda^2}\).
The mean \(mu\) of a continuous probability distribution p(x) is defined as:
\[ \mu = E(x) = \int x p(x) dx\]
For the exponential distribution, this can be integrated by parts, e.g., \(\int u dv = uv - \int v du\)
\[ \int ^\infty _0 x \lambda e^{-\lambda x} = -x e^{-\lambda x} - \int e^{-\lambda x} = -e^{-\lambda x}(\frac{1}{\lambda}+x)|^\infty _0\]
Using the limit of exp(-x) = 0 as x approaches infinity:
\[-\frac{e^{-\lambda x}}{\lambda}(1+x)|^\infty _0 = 0 + (\frac{1}{\lambda}+0) = \frac{1}{\lambda} \]
The variance $^2| of a continuous probability distribution p(x) is defined as:
\[ Var(x) = E[x^2] = (\int x^2 f(x) dx) - \mu^2\]
where \(\mu\) is the mean. Substituting in for the exponential distribution:
\[ \int ^\infty _0 x^2 \lambda e^{-\lambda x} dx\]
Using integration by parts again:
\[ \lambda \int ^\infty _0 x^2 e^{-\lambda x} dx = -x^2 e^{-\lambda x} + 2\int x e^{-\lambda x} dx = -x^2 e^{-\lambda x} - \frac{2}{\lambda}e^{-\lambda x}(\frac{1}{\lambda}+x)|^\infty _0\]
\[-x^2 e^{-\lambda x}+ 2\frac{e^{-\lambda x}}{\lambda}(1+x)|^\infty _0 = 0 + 0 + 0 +\frac{2}{\lambda} *\frac{1} {\lambda} = \frac{2}{\lambda^2} \]
\[ \frac{2}{\lambda^2} - [\frac{1}{\lambda}]^2 = \frac{1}{\lambda ^2}\]
The variance is equal to \(\frac{1}{\lambda}\).
Ex. 3 – It is estimated that there is a typo in every 250 data entries in a database, assuming the number of typos can obey the Poisson distribution. For a given 1000 data entries, what is the probability that there are exactly 4 typos? What is the probability of no typo at all? Use R to draw 1000 samples with \(\lambda = 4\) and show their histogram.
The Poisson distribution is \(p(x) \frac{\lambda^x e^{-x}}{x!}\), with x is the number of events we’re interested in observing, and \(\lambda = np\), where n is the total number of occurences and p is the probability of an event. In this example, p = \(\frac{1}{250}\), n = 1000 samples, and x = 4 typos. The exact solution is:
n = 1000
p = 1/250
x = 4
ex_3a <- ((n * p)^x * exp(-n*p))/factorial(x)
print(ex_3a)
## [1] 0.1953668
Approximately 0.195 of exactly 4 typos.
The probability of no typos is:
n = 1000
p = 1/250
x = 0
ex_3b <- ((n * p)^x * exp(-n*p))/factorial(x)
print(ex_3b)
## [1] 0.01831564
Approximately 0.018.
Using a Poisson random variable, below is a histogram of 1000 samples of \(\lambda = 1000 * \frac{1}{250} = 4\)
set.seed(123)
hist(rpois(n = 1000, lambda = 1000/250))
The most frequent outcome is approximately four typos.