Segment 28 Sanmit Narvekar

From Computational Statistics Course Wiki
Jump to navigation Jump to search

Segment 28

To Calculate

1. Draw a sample of 100 points from the uniform distribution . This is your data set. Fit GMM models to your sample (now considered as being on the interval ) with increasing numbers of components , at least . Plot your models. Do they get better as increases? Did you try multiple starting values to find the best (hopefully globally best) solutions for each ?

Here is the MATLAB code to fit GMMs to 100 samples drawn from the uniform distribution:

close all;

data = rand(100,1);

% Chose number of mixtures
K = 2;

% Initialize models "randomly"
mu = randsample(data, K)';
sigma = ones(1, K) * 0.3;

for iter=1:10

    pr = @(x) exp(-0.5*((x-mu)./sigma).^2)./(2.506 * sigma);
    prn = @(x) pr(x) ./ sum(pr(x));

    % E-step (compute assignments of all points)
    prns = zeros([numel(data),K]);
    for j=1:numel(data); prns(j,:)=prn(data(j)); end;

    % M-step

    % Re-estimate mean
    prnsx = zeros([numel(data), K]);
    for j=1:numel(data)
        prnsx(j,:) = data(j) * prns(j,:);
    mu = sum(prnsx) ./ sum(prns);

    % Re-estimate cov
    xmmu = zeros([numel(data), K]);
    for j=1:numel(data)
        xmmu(j,:) = data(j) - mu;
    sigma = sqrt(sum(prns .* (xmmu.^2)) ./ sum(prns));

    % Re-estimate P(k)
    Phat = sum(prns, 1) / numel(data);

    % Plot mixture pdf
    mixFunc = @(x) sum(Phat .* pr(x), 2);
    x = -1:.01:2;
    px = arrayfun(mixFunc,x);
    plot(x, px)


hold on;

[f x] = ksdensity(data);
hold off;
title(sprintf('K = %d', K))
legend('GMM', 'KS Density')

Here are the resulting mixture pdfs. In general, adding more mixtures does help. However, since the underlying distribution is NOT a mixture of Gaussians, it fails to capture the underlying distribution correctly.






2. Multiplying a lot of individual likelihoods will often underflow. (a) On average, how many values drawn from can you multiply before the product underflows to zero? (b) What, analytically, is the distribution of the sum of independent values , where ? (c) Is your answer to (a) consistent with your answer to (b)?

Part A

Here is the MATLAB code to compute it:

dist = zeros(10000,1);

for d=1:length(dist)
    a = 1;
    count = 0;
    while a ~= 0

        a = a * rand();
        count = count + 1;
    dist(d) = count;

mu = mean(dist)
std = std(dist)

On average, it takes 746.6751 +/- 27.2855 numbers before it underflows.

Part B

First we calculate the distribution of a single sample from log(U). The mean is as follows:

Recall the following property of expectation, where X is distributed according to f, and g is a function of X:

If we allow g to be log, and f to be the uniform distribution, then the mean becomes:

Then to calculate the variance, we first compute (using the same property as before):

Then the variance is:

Now, if we take the sum of N independent samples from log(U), they will be distributed as:

By the Central Limit Theorem, we can say they are distributed Normally with mean -N and variance N.

Part C

In part A, we found we could multiply about 746 numbers before reaching an underflow. Since this is equivalent to taking the log of the exp of these numbers, we can use our result from part B, which states that the mean of 746 numbers distributed as log(U) will be -746. Thus, we just check if the exp of -746 is the point at which we underflow.

MATLAB code:

for N=-740:-1:-750
    fprintf('%d %e\n', N, exp(N))

And the output, which is what we expect:

-740 4.199558e-322
-741 1.531604e-322
-742 5.434722e-323
-743 1.976263e-323
-744 9.881313e-324
-745 4.940656e-324
-746 0.000000e+00
-747 0.000000e+00
-748 0.000000e+00
-749 0.000000e+00
-750 0.000000e+00

To Think About

1. Suppose you want to approximate some analytically known function (whose integral is finite), as a sum of Gaussians with different centers and widths. You could pretend that (or some scaling of it) was a probability distribution, draw points from it and do the GMM thing to find the approximating Gaussians. Now take the limit , figure out how sums become integrals, and write down an iterative method for fitting Gaussians to a given . Does it work? (You can assume that well-defined definite integrals can be done numerically.)