{ The method I used to generate this: I took a plastic bowl and 10 coins, then for each round I threw the coins into the bowl, shook them (without looking, just in case), then rapidly turned it around and smashed it against the ground. I took the bowl up and wrote the ten generated bits by reading the coins kind of from "top left to bottom right" (heads being 1, tails 0). ~drummyfish }
Let's now take a look at how random the sequence looks, i.e. basically how likely it is that by generating random numbers by tossing a coin will give us a sequence with statistical properties (such as the ratio of 1s and 0s) that our obtained sequence has.
There are **494 1s and 506 0s**, i.e. the ratio is approximately 0.976, deviating from 1.0 (the value that infinitely many coin tosses should converge to) by only 0.024. We can use the [binomial distribution](binomial_distribution.md) to calculate the "rarity" of getting this deviation or higher one; here we get about 0.728, i.e. a pretty high probability, meaning that if we perform 1000 coin tosses like the one we did, we may expect to get the deviation we got or higher in more than 70% of cases (if on the other hand we only got e.g. 460 1s, this probability would be only 0.005, suggesting the coins we used weren't fair). If we take a look at how the ratio (rounded to two fractional digits) evolves after each round of performing additional 10 coin tosses, we see it gets pretty close to 1 after only about 60 tosses and stabilizes quite nicely after about 100 tosses: 0.67, 0.54, 0.67, 0.90, 0.92, 1.00, 0.94, 0.90, 0.88, 1.00, 1.04, 1.03, 0.97, 1.00, 0.97, 1.03, 1.10, 1.02, 0.98, 0.96, 1.02, 1.02, 1.02, 1.00, 0.95, 0.95, 0.99, 0.99, 0.99, 0.97, 0.95, 0.95, 0.96, 0.93, 0.90, 0.88, 0.90, 0.93, 0.95, 0.98, 0.98, 0.97, 0.97, 0.99, 1.00, 0.98, 0.98, 0.98, 0.97, 0.96, 0.95, 0.94, 0.95, 0.95, 0.96, 0.95, 0.96, 0.95, 0.96, 0.95, 0.96, 0.95, 0.96, 0.96, 0.97, 0.97, 0.97, 0.95, 0.94, 0.93, 0.93, 0.93, 0.94, 0.94, 0.94, 0.96, 0.95, 0.96, 0.96, 0.95, 0.96, 0.95, 0.95, 0.96, 0.97, 0.97, 0.96, 0.96, 0.95, 0.95, 0.95, 0.96, 0.97, 0.97, 0.97, 0.97, 0.96, 0.97, 0.98, 0.98.
Let's try the [chi-squared test](chi_squared_test.md) (the kind of basic "randomness" test): *D = (494 - 500)^2 / 500 + (506 - 500)^2 / 500 = 0.144*; now in the table for the chi square distribution for 1 degree of freedom (i.e. two categories, 0 and 1, minus one) we see this value of *D* falls somewhere around 30%, which is not super low but not very high either, so we can see the test doesn't invalidate the hypothesis that we got numbers from a uniform random number generator. { I did this according to Knuth's *Art of Computer Programming* where he performed a test with dice and arrived at a number between 25% and 50% which he interpreted in the same way. For a scientific paper such confidence would of course be unacceptable because there we try to "prove" the validity of our hypothesis. Here we put much lower confidence level as we're only trying not fail the test. To get a better confidence we'd probably have to perform many more than 1000 tosses. ~drummyfish }
We can try to convert this to a sequence of integers of different binary sizes and just "intuitively" see if the sequences still looks random, i.e. if there are no patterns such as e.g. the numbers only being odd or the histograms of the sequences being too unbalanced, we could also possibly repeat the chi-squared test etc.
The sequence as 100 10 bit integers (numbers from 0 to 1023) is:
Another way to test data randomness may be by **trying to [compress](compression.md) it**, since compression is basically based on removing regularities, redundancy, leaving only randomness. A compression algorithm exploits [correlations](correlation.md) in input data and removes that which can later be reasoned out from what's left, but with a completely random data nothing should be correlated, it shouldn't be possible to reason out parts of such data from other parts of that data, hence compression can remove nothing and it shouldn't generally be possible to compress completely random data (though of course there exists a non-zero probability that in rare cases random data will have regular structure and we will be able to compress it). Let us try to perform this test with the `lz4` compression utility -- we convert our 1000 random bits to 125 random bytes and try to compress them. Then we will try to compress another sequence of 125 bytes, this time a non-random one -- a repeated alphabet in ASCII (`abcdefghijklmnopqrstuvwxyzabcdef...`). Here are the results:
| sequence (125 bytes) | compressed size |
| -------------------- | ---------------- |
| our random bits | 144 (115.20%) |
| `abcdef...` | 56 (44.80%) |
We see that while the algorithm was able to compress the non-random sequence to less than a half of the original size, it wasn't able to compress our data, it actually made it bigger! This suggests the data is truly random. Of course it would be good to test multiple compression algorithms and see if any one of them finds some regularity in the data, but the general idea has been presented.
