Category: statistics

tl;dr

I needed to generate Standard Normal Random Variates for a simulation I’m programming, and I wanted to see which method would be best to use. There are a few options, but generally you need to experiment to know for sure which will work best in your situation.

I decided to do this in Python, and to share my code + experimentation process and recommendations/reasoning.

Hoping this can save someone else some time later. If you need to do the same for your own simulation, or any other reason, hopefully this solves some of your problems, helps you understand what’s going on under the hood, and just makes life easier.

Grab the code here:  https://github.com/adpoe/STD-Normal-Rand-Variates/tree/master/code

Generating Random Variates from Standard Normal

Experimental Setup

I used Python to algorithmically generate random variates that follow the Standard Normal distribution according to three different methods. For all methods, 10,000 valid random variables were generated in each algorithm’s run, in order to maintain consistency for later effectiveness comparisons. The methods tested were:

1. The Inverse Transform Method
2. The Accept/Reject Method
3. The Polar-Coordinates Method

In the following paragraphs, I will briefly outline the implementation decisions made to generate Standard Normal random variates according to each method. I will then analyze the results, compare them, and issue my own recommendation on which method to use going forward, informed by the data gathered in this experiment.

Inverse Transform

The Inverse Transform Method works by finding the inverse of the CDF for a given probability distribution (F^-1(X)), then feeding random numbers generated from U[0,1) into that inverse function. This will yield randomly generated variables within the range of our desired probability distribution.

However, this method is problematic for the Standard Normal Distribution, because there is no closed form for its CDF, and hence we cannot calculate its exact inverse. Because of this, I chose to use Bowling’s closed-form approximation of the Standard Normal CDF, which was developed in 2009: Pr(Z <= z) = 1 / [1 + e^(-1.702z)].

Despite being only an approximation, Bowling’s closed form CDF function is mathematically close enough to generate reasonable random variates. Beyond that, this function is simple. The hardest part was calculating the inverse, which was actually done with help from Wolfram Alpha. Once an inverse was obtained, implementation was straightforward and can be seen in the code attached, within the method @inverse_transform().

Accept/Reject

The Accept/Reject method for random variate is more complex, and it can be implemented a few different ways. I chose to use the method outlined by Sheldon Ross in his book Simulation (Fifth Edition), on page 78.

The procedure, and a snippet of the core code used, are both presented in-line below, as an illustration:

To note: Our Y1 and Y2 values are seeded with variates generated by the exponential distribution, with lambda=1. In much of the literature on the Accept/Reject method, this function is called “g(x)”, and for this example we used the exponential distribution.

It is also important to keep track of how many rejections we get while using the Accept/Reject method. In order to determine the average number of rejections, I ran the algorithm 10 times. The data set it created is shown below:

While generating 10,000 variates, the algorithm created 3130.8 rejections, on average. This means that, generally, there was about 1 rejected variate for every 3 valid variates.

Polar-Coordinate Method

In the Polar-Coordinate method, we take advantage of trigonometric properties by generating random variables uniformly distributed over (0, 2pi), and then transforming them into rectangular coordinates. This method, called the Box-Muller Method more formally, is not computationally efficient, however, because it involves use of logs, sines, and cosines—all expensive operations on a CPU.

In order to perform this operation more efficiently, I’ve used the method outlined by Sheldon Ross, on page 83 of his book Simulation (5^th Ed.).

Step 1: Generate random numbers, U₁ and U₂

Step 2: Set V₁ = 2U₁ – 1, V₂ = 2U₂ – 1, S = V₁² + V₂²

Step 3: If S > 1, return to Step 1.

Step 4: Return the independent standard normal for two variables, X and Y:

Where:

X = sqrt(-2*log(S)/2) * V₁,

Y = sqrt(-2*log(S)/S)*V₂

Expectations

Prior to running the experiment, I expected the Inverse-Transform Method to generate the worst variables themselves, because it only uses an approximation of the Standard Normal CDF, not the CDF itself. I was a little nervous about using an approximation for a CDF with no closed form to generate my inverse function, thinking that while our inverse may deliver results that are more or less reasonable, the resulting data set wouldn’t pass more advanced statistical tests since we are presumably losing precision, through the approximation process. But that said, I also expected its time efficiency to be the best, because we are only calculating logarithm each time we call the inverse function, and this seems to be the only slow operation.

I expected that method 2, Accept/Reject generate the most accurate variables, mostly because of the convincing mathematical proofs describing its validity on pages 77 and 78 of Ross’s Simulation textbook. Intuitively, the proof for this method makes sense, so I expected its data set to look most like something that truly follows the Standard Normal Distribution. From a time efficiency standpoint however, I expected this algorithm to perform 2^nd best, because I’m using a logarithm each time I generate an exponential random variable. And with 2 log calls for each run, it seems like this method would be relatively slow, under the assumption that Python’s log function is expensive. (Log calls are used here because we know that –logU is exponential with rate lambda=1. But we need exponential variables generated with rate 1 for each Y variable, Y₁ and Y₂.)

The Polar Coordinate Method is the most abstract for me, and so I had a hard time seeing exactly why it would generate Standard Normal Random variables, and because of this, I wasn’t sure what to expect of its data set. I took it on faith that it would generate the correct variables, but I didn’t fully understand why. Moreover, I also expected it to perform the worst from a runtime perspective because it involves the most expensive operations: Two Square Roots and Two Log calls for each successful run.

Histograms and Analysis

In order to determine whether each method produced qualitatively accurate data, I then plotted histograms for the numbers generated by each. All three histograms can be seen below. On my examination, it appears that the Inverse Transform yielded the least accurate representation, while the Polar Coordinate Method and Accept/Reject were much better, and about equal in validity.

Notably, the Inverse Transform method generated many values beyond the expected range of the Standard Normal (greater than 4.0 and less than -4.0). And while these values are possible, it seems like too many outliers—more than we would see in a true Standard Normal Distribution. Again, I think this is because we are using an approximation for the CDF, rather than the true Inverse of the CDF itself. I had some trouble getting these graphs to sit in-line, so please review the appendix to see them.

Chi-Squared Test

In order to determine wither the numbers generated may be from the Normal Distribution, I ran each method 10 times, and performed a Chi-Square test on each result. The data set can be seen in the tables within the appendix at the back of this document.

From this test, I was able to make the following determinations:

• Inverse Transform:
  • N=10,000
  • Avg Chi Sq: 2,806.719
    • From 10 tests, each n=10,000
  • Critical Value:749
  • Result:  REJECT Null Hypothesis

• Accept/Reject
  • N=10,000
  • Avg Chi Sq: 10,025.226
    • From 10 tests, each n=10,000
  • Critical Value: 10,233.749
  • Result:  ACCEPT Null Hypothesis

• Accept/Reject
  • N=10,000
  • Avg Chi Sq:   9,963.320
    • From 10 tests, each n=10,000
  • Critical Value: 10,233.749
  • Result:  ACCEPT Null Hypothesis

Runtime Analysis

Again, I ran each method 10 times to collect a sample of data with which to analyze runtime.

The average runtimes from 10 tests with n=10,000 were:

• Inverse Transform: -6.60E-06 seconds
• Accept/Reject: -5.72E-06 seconds
• Polar Coordinates: –63E-06 seconds

This result was indeed surprising. I had expected the Polar Coordinates method to perform the worst, and it did—but only by a very small margin. Moreover, I had expected Inverse Transform to perform the best, and it was only fractions of a microsecond (or nanosecond?) faster than Polar Coordinates on average. I did not expected Accept/Reject to perform so well, but it was by far the fastest overall.

Conclusion

Given these results, I would recommend the Accept/Reject method for anyone who needs to generate Standard Normal Random Variables in Python (at least using my same implementations).

Accept/Reject not only passed the Chi-Square test at the 95% significance level, it also was by far the fastest of the three algorithms. It is roughly comparable to Polar Coordinates on the quality of variables generated, and it beats everything else on speed.

APPENDIX – FIGURES:

Fig 1.1 – Inverse Transform Method Histogram

Fig 1.2 – Polar Coordinates Method HistogramFig 1.3 – Accept/Reject Method Histogram

DATA ANALYSIS TABLES

tl;dr

An example of what you’ll find:

I generated some random numbers with a few different generators, some of which I made, and also used the one provided directly by Python. I wanted to see what the best option is for generating random numbers in a few simulations I’m doing.

Grab code here: >> lcg.py << 

It’s commented and can be run by simply invoking Python with:  “python lcg.py”

I tried to explain what I was doing at each step to make this clear even for the comparatively un-initiated to the more esoteric statistics at play here, which aren’t totally necessary to know, and really will just be an impediment to __getting_started_now__. I wanted to share something so that people who are practically-minded like me can just jump in, start messing around, and know what to expect. So hopefully with just this code (full repo linked below) and the information presented here, you can start messing around with random number generators if you need to for any reason. (Most common reason would be to seed random variates in a simulation.)

My main goal in posting this is to give anyone with an interest in generating randomness an easy entry into it–with working code for these sort of generators, as it’s somewhat hard to find online, and the details can be a bit opaque, without clear examples of what to expect when you’re testing. So ideally, this will be the total package. Probably not, but hey, giving it a try. If for some reason you need help, feel free to contact me.

Random Number Generation

 Generating truly random numbers is a longstanding problem in math, statistics, and computer science. True randomness requires true entropy, and in many applications—such as generating very large sets of random numbers very quickly—sufficient “true” entropy is difficult or impractical to obtain. (Often, it needs to come from the physical environment, sources such as radioactive decay, etc.) So, instead, we look to algorithmic random number generators for help.

Algorithmically generated random numbers will never be “truly” random precisely because they are generated with a repeatable algorithmic formula. But for purposes such as simulating random events – these “Pseudo-random” numbers can be sufficient. These algorithmic generators take a “seed value” from the environment, or from a user, and use this seed as a variable in their formula to generate as many random-like numbers as a user would like.

Naturally, some of these algorithms are better than others, and hundreds (if not thousands, or more) of them have been designed over the years. The output is always deterministic, and never “truly” random, but the ideal goal is to approximate randomness by generating numbers which:

1. Are uniformly distributed on the range of [0,1)
2. Are statistically independent of each other
   1. (That is, the outcomes of any given sequence do not rely on previously generated numbers)

The best random number generators will pass statistical tests for both uniformity and independence. In this analysis, we will subject three different random number generation algorithms to series of statistical tests and compare the outcomes.

The algorithms we will test are:

• Python’s Built-In Random Number Generator
  • This algorithm is called the “Mersenne Twister”, implementation details are available at: Python Docs for Random
  • Seed value: 123456789
• A Linear Congruential Generator
  • Seed value: 123456789
  • a=101427
  • c=21
  • m=2¹⁶
• A Linear Congruential Generator with RANDU initial settings
  • Seed value: 123456789
  • a=65539
  • c=0
  • m=2³¹

The tests each algorithm will be subjected to are:

• Uniformity tests
  • Chi-squared Test for Uniformity
  • Kolmogorov-Smirnov Test for Uniformity
    • Null hypothesis for BOTH tests: The numbers in our data set are uniformly distributed
  • Independence tests
    • Runs Test for Independence
    • Autocorrelation Test for Independence, (gap sizes: 2,3,5, and 5 will be used)
      • Null hypothesis for BOTH tests: The numbers in our data set are independent of each other

Test Result Summary & Quick Notes on Implementation 

The exact implementation of each test can be viewed in the linked Python file named: “lcg.py”. The tests can be duplicated by anyone with Python installed on their system by running the command “python lcg.py”.

You can view the file directly on GitHub here:  >> lcg.py <<

SUMMARY TABLE (2.1)

The summary table above shows each algorithm tested, and which tests were passed or failed. More detailed output for each test and for each algorithm can be viewed in Tables 1.1 – 1.3 in the appendix to this document.

The Kolmogorov-Smirnov (or KS test) was run at the following levels of significance: .90, 0.95, 0.99. The formulas for the critical value at these significance levels were taken from table of A7 of Discrete-Event System Simulation by Jerry Banks and John S. Carson II. (Formulas for 0.80 could not be found, so I’ve used what was available.)

All other tests were run at the 0.80, 0.90., and 0.95 significance level.

Statistical Analysis and Expectations 

Prior to generating the numbers for each test, I expected Python’s random function to perform the best of all three algorithms tested, mostly because it’s the library random function of one of the world’s most popular programming languages. (Which means: thousands and thousands of code repositories rely on it—many of which are used by commercial and mission critical programs.) But in fact, it performed the worst, failing the Runs Test at both the 0.80 and 0.90 level of significance. These failings are NOT statistically significant at the alpha=0.05 level, but it’s still surprising to see.

I expected the RANDU algorithm to perform the worst, and I thought it would perform especially badly on the autocorrelation test. This is because RANDU is known to have problems, outlined here. Specifically, it is known to produce values which fall along only a specific set of parallel planes (visualization in link above), which means the numbers should NOT be independent, when tested at the right gap lengths. It’s possible that the gap lengths I’ve tested simply missed any of these planes, and as a result—RANDU performed the best of all the algorithms. It’s the only algorithm that didn’t fail any statistical tests at all.

I anticipated the LGC function to perform 2^nd best overall, and I was right about that—but the best and worst algorithm were the opposite of what I expected. Mostly, I thought that that Python’s random generator would be nearly perfect, RANDU would be badly flawed, and the LGC would be just okay. Really, the LGC performed admirably: The only test it failed was autocorrelation at the 0.80 confidence level, and that isn’t statistically significant by most measures.

Reviewing the data output into each .txt file directly, I don’t see any discernible patterns in the numbers themselves. And with 10,000 data points, there’s so much output to review that I can see why statistical measures are needed to effectively to determine what’s really going on in the data. It’s probably possible to find a few patterns, specifically related to runs and gap-sequences just by viewing the data directly, but tests are still needed to find out for sure.

With that said, I do think the testing done in this experiment is sufficient, because we have two tests for each measure that matters: 1) Uniformity; 2) Independence. The only improvement I would make for future tests is testing more gap-sequences, and starting them at different points. I’d do this mostly because I know that RANDU should fail gap-sequence tests given the right input, but there would be some trial and error involved in trying to find these sequences naively. So, sometimes, getting into math itself and working with proofs may still be the most effective method. Maybe sometime the old-fashioned way is still best.

TABLE 1.2 – Linear Congruential Generator (X₀ = 123456789 )

END TABLE 1.2

TABLE 1.3 – Linear Congruential Generator with RANDU initial settings 

END TABLE 1.3

Table 2.1 – Test Result Comparisons