Difference between revisions of "Likelihood"

Revision as of 16:23, 26 January 2009

The likelihood tells you how probable it is that your model produces your data. This measure tells you how good your model is. First we need to understand the concept of likelihood, and this brings us into probability theory. We start by studying random numbers on a computer.

Random Numbers in MATLAB

Start MATLAB and type the following into the MATLAB prompt

rand

This produces a random number. Repeat several times. Continue typing the material in boxes into the MATLAB prompt.

help rand

Returns help for the rand command. We'll discuss this help file. Try

rand(3)
rand(3,2)
rand(3,2,2)

These commands allow you to create arrays of random numbers without loops. Compare how long it takes to run

A = rand(10000,1);

With

for i = 1:10000
B(i,1) = rand;
end

Note however there is really two issues with the second method: first, we could have avoided the loop, and second, the we successively built a larger and larger array without first initializing it. Now that B is defined, repeating the second set of command will go much faster, but still not as fast as avoiding the loop.

Note semicolons suppress a command's insistence on displaying the result on the screen. Try it without the semicolon:

A = rand(10000,1)

Finally, if you are running a model that uses random numbers it is useful to be able to get the same random numbers again to reproduce your results. This can be done: computer random numbers are not really random, they are pseudo-random. The number returned by the generator is completely determined by the "seed" used to initialize the generator and the number of times that the generator has been used since the last initialization. Try:

rand('state',0);
rand
rand
rand
rand('state',0);
rand
rand
rand
rand('state',1);
rand
rand
rand

The command "clock" can be used to set the seed in "random" way. Use "up arrow" to repeat a command.

clock
clock
clock
sum(100*clock)
sum(100*clock)
sum(100*clock)
rand('state',sum(100*clock));
rand
rand
rand('state',sum(100*clock));
rand
rand

If you want a "random" seed and still be able to reproduce your simulations use:

s = sum(100*clock);
rand('state',s);
rand
rand
rand('state',s)
rand
rand

Probability

We want to use the command rand to roll a die. The command rand returns a number between 0 and 1. So 6*rand returns a number between 0 and 6. The command ceil rounds numbers up.

help ceil

ceil(6*rand)

Repeat this command several times. But now let's create a new command. Open the Editor and type

function d = die
% DIE Roll a die
% d = die returns a random integer between 1 and 6, inclusive

d = ceil(6*rand);

Save the file as die.m then type

die
die
die

For the next step it is useful to get more than one throw of the die. Save as dice.m then change

function d = dice(n)
% DICE Rolls dice
% d = dice(n) returns n random integers between 1 and 6, inclusive

d = ceil(6*rand(n,1));

Now some statistics:

mean(dice(100))
mean(dice(1000))
mean(dice(10000))
mean(dice(100000))

• Homework Problem 1: Roll the dice many times. Make a guess as to whether the mean converges as the number of dice increases. Assuming you do not know the answer, what numerical experiments would increase your confidence in this regard?

Now try the same thing for

median
mode
std
min
max

The command std is the standard deviation.

• Homework Problem 2: Roll the dice many times. Make guesses as to whether the statistics median, mode, std, min and max converge as the number of dice increases. For the ones that converge, which limits are equal.

Probability Mass Function

The PMF gives the theoretical fraction of times each value is returned on repeated experiments. For a fair die

f(i) = 1/6, for i = 1,2,3,4,5,6

f(i) = 0 otherwise.

For a loaded die, some the six non-zero values of f would differ, but they must add to 1.

The statistics considered above can be calculated theoretically from the PMF. Then there are theorems that show convergence of these statistics to their theoretical values ... or counterexamples showing that convergence does not happen. Warning: Convergence is a subtle concept in Probability Theory.

Of note: The theoretical "mode" for loaded die is the i which maximizes f. For a fair die, all values are equal and the "mode" does not exist, theoretically. The idea of "most probable" is important to us.

Plotting the PMF: Roll the dice and create a histogram with 6 bins:

help hist
type hist

A = dice(100);
hist(A,6)

Repeat several times. Then try:

A = dice(1000);
hist(A,6)

A = hist(10000);
hist(A,6)
hist(A,60)
hist(A,600)

Notice that as the width of the bin gets smaller, the number in each bin (containing the integers 1,2,3,4,5,6) does not change. Also notice that the height of each bin, divided by the total number of dice, equals

1/6

Probability theory defines probability with axioms (we will not study the axioms). The most common interpretation of probability of an outcome is that it is the fraction of times it occurs when if the experiment (roll the die) could be repeated many times. The PMF gives the probability of occurrence of each discrete outcome: 1,2,3,4,5, or 6.

Probability Density Function

Probabilists make an analogy between probability and mass. Mathematically, you can think of probability (or mass) as concentrated at discrete mathematical points (1,2,3,4,5,6) or spread out across a continuum. Think of a continuous solid: a 10 kilo barbell. How much mass is there in 1 cubic centimeter of the volume of the barbell? Much less than 10 kilos! How much mass in 1 cubic millimeter? The mass in 1 cubic millimeter is less than the mass in one cubic centimeter by a factor of 1000! As you take smaller and smaller volumes you get less and less mass. Ignoring atoms and elementary particle, suppose the barbell is continuous solid. The smaller the volume the less mass in that volume. And there is no mass at a mathematical point! But the ratio of mass to volume typically converges: it is called the density. If mass were concentrated at individual points (as with the dice) the density at those points would be infinite!

Now lets compare dice and rand. Dice returned a discrete number: 1,2,3,4,5,6. Rand returns a "continuous" number, or approximately continuous as nothing on a digital computer is really continuous, if you zoom way in. But we won't zoom in. For dice, the number of occurrences in each bin (containing exactly one of the numbers 1,2,3,4,5, or 6) did not depend on the size of the bin. Now try:

A = rand(1000,1);
hist(A,10)
hist(A,100)
hist(A,1000)
hist(A,10000)

Here is some terminology: Dice and rand return what is called a random variable, which is some function of the state of the random number generator. (The mathematical definition of random variable is similar.) Dice returns a discrete random variable rand returns a "continuous" random variable. Discrete random variables have probability mass functions (PMFs) the probability density at points with nonzero mass is infinite. Continuous random variables have probability density functions (PDFs) with zero probability mass everywhere. There is such a thing as a mixed random variable which is part continuous and part discrete, but we will not get into it.

You can visualize PMFs and PDFs both with the hist command, although the vertical scale will be wrong.

A = rand(100000,1);
hist(A,100)

A different random number generator with a different PDF: randn; "n" for normal.

A = randn(100000,1);
hist(A,100)

Note the famous bell curve.

Some facts about PDFs, densities, and the bins from the command hist:

• The density at a point is approximately the fraction of points that fall into its bin divided by the width of the bin.
• Don't like approximations? Think limit.
• The probability of landing in a bin is approximately the density at one point in the bin times the width of the bin.
• Don't like approximations? Think integral.
• For discrete random variables, the probability masses (nonzero values of PMF) must add to 1.
• For continuous random variables, the area (integral) under the PDF must be 1.
• Probability mass must be less than or equal to 1; probability densities can approach infinity.

Functions of Random Variables

We have two random number generator (rand and randn). In MATLAB that's basically all you have. You can get more interesting random variables by taking function of random variables. Remember the die

ceil(6*rand)

Let's define some more functions. Open affine.m in the editor and type in:

function D = affine(A,m,b)
% AFFINE affine transformation of data D

D = m*A + b;

Save. Now open sigmoid.m and type in

function D = sigmoid(A,k,h)
% SIGMOID sigmoid transformation of data D

D = 1/(1+exp(-(A-h)/k));