Skip to main content

A personal annotations of Veach's thesis (10) P.60

Star discrepancy D_N^*



I might say, the goal of Quasi-Monte Carlo method seems to have a sort of contradictory. It is:
  1. To minimize the irregularity of distribution
  2. Not using regular uniform samples to avoid aliasing 
To minimize the irregularity of distribution, a straightforward answer is using regular distributed sampling. But, this causes aliasing that is the point to use the Monte-Carlo method (Note: not Quasi-Monte-Carlo).

A discrepancy is a quantitative measure of irregularity of distribution. This measure tells us how much integration error occurs.

The star discrepancy D_N^* is a discrepancy. When we sample many boxes that always includes the origin, it shows the maximum error of estimation of the area of these boxes.

In the case of one dimension, boxes are lines. Therefore we estimate the length of line that includes the origin by sampling. If we sample N points, if we sample regular uniformly, we could not sample less than 1/(2N) length correctly. This is the same as the sampling theory. But, if we go to the high dimensional area (s-dimensional area), the error bound becomes O((log(N)^s)/N).

I don't understand why it is O((log(N)^s)/N). It is s-dimensional generalization, but, it even doesn't work one dimensional case. Some told me there is a lengthy complicated proof. Figure 1 shows the plot. When number of samples are small and dimension is high, then the error becomes larger. This is the growth ratio of log(N)^s and N, someone says it is obvious, but I would like to see that anyway.

This is a program to show this plot. This is matlab/octave language.

%
% visualize star discrepancy
% Octave code: 2010 (C) Shitohichi Umaya
%
x = [10000:100:1000000];
s = 2
y1 = power(log(x), s)./x;
s = 3
y2 = power(log(x), s)./x;
plot(x,y1,"1;(log N)^s/N, s=2;", x,y2,"3;(log N)^s/N, s=3;")

Acknowledgements:
Thanks to Alexander K. to answer me what is the discrepancy.

Comments

Popular posts from this blog

Why A^{T}A is invertible? (2) Linear Algebra

Why A^{T}A has the inverse Let me explain why A^{T}A has the inverse, if the columns of A are independent. First, if a matrix is n by n, and all the columns are independent, then this is a square full rank matrix. Therefore, there is the inverse. So, the problem is when A is a m by n, rectangle matrix.  Strang's explanation is based on null space. Null space and column space are the fundamental of the linear algebra. This explanation is simple and clear. However, when I was a University student, I did not recall the explanation of the null space in my linear algebra class. Maybe I was careless. I regret that... Explanation based on null space This explanation is based on Strang's book. Column space and null space are the main characters. Let's start with this explanation. Assume  x  where x is in the null space of A .  The matrices ( A^{T} A ) and A share the null space as the following: This means, if x is in the null space of A , x is also in the n...

Gauss's quote for positive, negative, and imaginary number

Recently I watched the following great videos about imaginary numbers by Welch Labs. https://youtu.be/T647CGsuOVU?list=PLiaHhY2iBX9g6KIvZ_703G3KJXapKkNaF I like this article about naming of math by Kalid Azad. https://betterexplained.com/articles/learning-tip-idea-name/ Both articles mentioned about Gauss, who suggested to use other names of positive, negative, and imaginary numbers. Gauss wrote these names are wrong and that is one of the reason people didn't get why negative times negative is positive, or, pure positive imaginary times pure positive imaginary is negative real number. I made a few videos about explaining why -1 * -1 = +1, too. Explanation: why -1 * -1 = +1 by pattern https://youtu.be/uD7JRdAzKP8 Explanation: why -1 * -1 = +1 by climbing a mountain https://youtu.be/uD7JRdAzKP8 But actually Gauss's insight is much powerful. The original is in the Gauß, Werke, Bd. 2, S. 178 . Hätte man +1, -1, √-1) nicht positiv, negative, imaginäre (oder gar um...

Why parallelogram area is |ad-bc|?

Here is my question. The area of parallelogram is the difference of these two rectangles (red rectangle - blue rectangle). This is not intuitive for me. If you also think it is not so intuitive, you might interested in my slides. I try to explain this for hight school students. Slides:  A bit intuitive (for me) explanation of area of parallelogram  (to my site, external link) .