# HTW Berlin - Angewandte Informatik - Advanced Topics - Exercise - Monte Carlo Estimator

## Introduction

In this notebook you will first see why it is important to draw samples according to the probability distribution and not just using a grid (np.linspace(...)) or a uniform distribution. Afterwards, you will implement inverse transform sampling and use it to estimate a function with respect to a probability density function. Finally, you will make use of control variates to reduce the variance of the estimator and to improve the estimation.

In order to detect errors in your own code, execute the notebook cells containing assert or assert_almost_equal. These statements raise exceptions, as long as the calculated result is not yet correct.

## Requirements

### Knowledge

To complete this exercise notebook, you should possess knowledge about the following topics.

• Probability density function
• Probability mass function
• Expected value (Exercise - Expected Value)
• Monte Carlo estimator
• Inverse transform sampling
• Variance reduction

### Python Modules

# External Modules
import matplotlib.pyplot as plt
import numpy as np
from numpy.testing import assert_almost_equal

%matplotlib inline
def plot_p_f_and_samples(x, p, f, samples_x1=[], samples_x2=[]):
"""
Plots two functions p(x) and f(x) for given x values
and f(samples_x) as single points

:param x: x-values for the functions p and f
:type x: 1D ndarray
:param p: pdf
:type p: lambda function
:param f: function evaluated with x and samples_x
:type f: lambda function
:param samples x: x-values for the function f drawn as single points
"""
plt.plot(x, p(x), label='p(x)')
plt.plot(x, f(x),label='f(x)')
if len(samples_x1) > 0:
plt.plot([samples_x1[:-1]], [f(samples_x1)[:-1]], marker='x', color='g')
plt.plot([samples_x1[-1]], [f(samples_x1)[-1]], marker='x', color='g', label='f(samples_x1)')
if len(samples_x2) > 0:
plt.plot([samples_x2[:-1]], [f(samples_x2)[:-1]], marker='x', color='r')
plt.plot([samples_x2[-1]], [f(samples_x2)[-1]], marker='x', color='r', label='f(samples_x2)')
plt.xlabel('x')
plt.legend(loc='upper left')

### Crude Monte Carlo Estimator (CMC)

We have a function, which estimates$\mathbb E_{x \sim p} [f(x)]$ with the use of$n$ drawn samples, with the given functions:

$f(x)=\sin(x)$

and $p(x) = \text{uniform}(0, \pi/2)$

### Exercise - Variance Reduction by Control Variates

Determine the empirical variance (also sample variance) of the CMC-estimator for n = 10 by computing the expected value 10,000 times.

# Complete this cell
#(...)

#(...)
e_variance = None
# Executing this cell must not throw an Exception
# The solution is obfuscated so you can solve the exercise without unintendedly spoiling yourself

#obfuscated_solution = 2366.531932907348 * 626 / 72727
#assert_almost_equal(e_variance, obfuscated_solution, decimal=2)

The following plot visualizes how the estimation converges towards the true expected value as n increases.

# PLOT

n=500
x_, fs_mean, fs = estimate_f_CMC(n=n)
x = np.arange(n-1)
e = np.ndarray(n-1)
for i in range(1,n):
e[i-1] = fs[:i].mean()

print(fs.mean())

plt.figure(figsize=(20,4))
plt.plot(x,e)
#plt.axhline(y=2., c='g')
plt.xlabel("num of samples")
plt.ylabel("estimate")

### Exercise - Variance Reduction by Control Variates

• Reduce the empirical variance and improve the estimator with the use of control variates.
• What's the error (empirical standard deviation) of the estimation?

Hint:

We have -$\hat F$ is CMC the estimator of$\mathbb E_p [f(x)]$: $\hat F = \frac{1}{n}\sum_i^n f(x^{(i)})$

Search for a variable$g(x)$, which fulfills the following:

*$g(x)$ is correlated with$f(x)$

• its expected value$\mathbb E[g(X)]$ is known

Reminder: $\bar F$ is the improved estimated value

$\bar F = \hat F - \beta \cdot ( G - \mathbb E[g(X)] )$

with $G = \frac{1}{n}\sum_i^n g(x^{(i)})$ with optimum$\beta$:

$\beta = \frac{cov(\hat F,\bar F)}{var(\bar F)}$

# Complete this cell

# get some values for V and W
x_, fs_mean, fs =  estimate_f_CMC(n=10)

# (...)

b = None# calculate beta first
# Executing this cell must not throw an Exception
# The solution is obfuscated, so you can solve the exercise without unintendedly spoiling yourself

#obfuscated_solution = TODO
#assert_almost_equal(b,obfuscated_solution)
# Now calculate the improved estimation of f

# (...)

Z = None
# Executing this cell must not throw an Exception

solution = .636
assert_almost_equal(cmc_estimates.mean(), solution, decimal=3)

## Literature

The following license applies to the complete notebook, including code cells. It does however not apply to any referenced external media (e.g., images).

HTW Berlin - Angewandte Informatik - Advanced Topics - Exercise - Monte Carlo Estimator
by Christian Herta, Klaus Strohmenger