Should this number be subtracted by 1 when implementing the Mersenne Twister? - javascript

I found this snippet online along with this Stackoverflow post which converts it into a TypeScript class.
I basically copy and pasted it verbatim (because I am not qualified to modify this sort of cryptographic code), but I noticed that VS Code has a little underline in the very last function:
/**
* generates a random number on [0,1) with 53-bit resolution
*/
nextNumber53(): number {
let a = this._nextInt32() >>> 5;
let b = this._nextInt32() >>> 6;
return (a * 67108864.0 + b) * (1.0 / 9007199254740992.0);
}
Specifically the 9007199254740992.0
VS Code says Numeric literals with absolute values equal to 2^53 or greater are too large to be represented accurately as integers.ts(80008)
I notice that if I subtract that number by one and instead make it 9007199254740991.0, then the warning goes away. But I don't necessarily want to modify the code and break it if this is indeed a significant difference.
Basically, I am unsure, because while my intuition says that having a numerical overflow is bad, my intuition also says that I shouldn't try to fix cryptographic code that was posted in several places, as it is probably correct.
But is it? Or should this number be subtracted by one?

9007199254740992 is the right value to use if you want Uniform values in [0,1), i.e. 0.0 <= x < 1.0.
This is just the automatics going awry, this value can be accurately represented by a JavaScript Number, i.e. a 64bit float. It's just 253 and binary IEEE 754 floats have no trouble with numbers of this form (it would even be represented accurately with a 32bit float).
Using 9007199254740991 would make the range [0,1], i.e. 0.0 <= x <= 1.0. Most libraries generate uniform values in [0,1) and other distributions are derived from that, but you are obviously free to do whatever is best for your application.
Note that the actual chance of getting the maximum value back is 2-53 (~1e-16) so you're unlikely not actually see it in practice.

Related

Will I possibly loose any decimal digits (precision) when multiplying Number.MAX_SAFE_INTEGER by Math.random()?

Will I possibly loose any decimal digits (precision) when multiplying Number.MAX_SAFE_INTEGER by Math.random() in JavaScript?
I presume I won't but it'd be nice to have a credible explanation as to why šŸ˜Ž
Edited, In layman terms, we're dealing with two IEEE 754 double-precision floating-point numbers, one is the maximal integer (for double-precision), the other one is fractional with quite a few digits after a decimal point. What if (say) I first converted them to quadruple-precision format, then multiplied, and then converted the product back to double-precision, would the result be any different?
const max = Number.MAX_SAFE_INTEGER;
const random = Math.random();
console.log(`\
MAX_SAFE_INTEGER: ${max}, \
random: ${random}, \
product: ${max * random}`);
For more elaborate examples, I use it to generate BigInt random numbers.
Your implementation should be safe - in theory, all numbers between 0 and MAX_SAFE_INTEGER should have a possibility of appearing, if the engine implementing Math.random uses a completely unbiased algorithm.
But an absolutely unbiased algorithm is not guaranteed by the specification - the numbers chosen are meant to be psuedo random, not truly, completely random. (does such a thing even exist? it's debatable...) Modern versions V8 and some other implementations use an algorithm with a period on the order of 2 ** 128, larger than MAX_SAFE_INTEGER (2 ** 53 - 1) - but it'd be completely plausible for other implementations (especially older ones) to have a much smaller period, resulting in certain integers within the range being picked much more often than others.
If this is important for your script (which is pretty unlikely in most situations, I'd think), you might consider using a higher-quality random generatior than Math.random - but it's almost certainly not worth worrying about.
What if (say) I first converted them to quadruple-precision format, then multiplied, and then converted the product back to double-precision, would the result be any different?
It could be in cases where the rounding behaves differently between multiplying two doubles vs converting quadruple to double, but the main problem remains the same. The spacing between representable doubles in the range from 2n to 2n+1 is 2nāˆ’52. So between 252 and 253 only whole numbers can be represented, between 251 and 252 only every 0.5 can be represented, etc.
If you want more precision you could try decimal.js. The library is included on that documentation page so you can try these out in your console.
Number.MAX_SAFE_INTEGER*.9
8106479329266892
new Decimal(Number.MAX_SAFE_INTEGER).mul(new Decimal(0.9)).toString()
"8106479329266891.9"
Both answers are correct, but I couldn't help running this little experiment in C#, where double is the same thing as Number in JavaScript (fiddle):
using System;
public class Program
{
public static void Main()
{
const double MAX_SAFE_INT = 9007199254740991;
Decimal maxD = Convert.ToDecimal(MAX_SAFE_INT.ToString());
var rng = new Random(Environment.TickCount);
for (var i = 0; i < 1000; i++)
{
double random = rng.NextDouble();
double product = MAX_SAFE_INT * random;
// converting via string to workaround the "15 significant digits" limitation for Decimal(Double)
Decimal randomD = Decimal.Parse(String.Format("{0:F18}", random));
Decimal productD = maxD * randomD;
double converted = Convert.ToDouble(productD);
if (Math.Floor(converted) != Math.Floor(product))
{
Console.WriteLine($"{maxD}, {randomD, 22}, products: decimal {productD, 32}, converted {converted, 20}, original {product, 20}");
}
}
}
}
As far as I'm concerned, I'm still getting the desired distribution of the random numbers within the 0 - 9007199254740991 range.
Here is a JavaScript playground code to check for possible recurrences.

Javascript scientific notation floats vs integers explained

I am working with js numbers and have lack of experience in that. So, I would like to ask few questions:
2.2932600144518896
e+160
is this float or integer number? If it's float how can I round it to two decimals (to get 2.29)? and if it's integer, I suppose it's very large number, and I have another problem than.
Thanks
Technically, as said in comments, this is a Number.
What you can do if you want the number (not its string representation):
var x = 2.2932600144518896e+160;
var magnitude = Math.floor(Math.log10(x)) + 1;
console.log(Math.round(x / Math.pow(10, magnitude - 3)) * Math.pow(10, magnitude - 3));
What's the problem with that? Floating point operation may not be precise, so some "number" different than 0 should appear.
To have this number really "rounded", you can only achieve it through string (than you can't make any operation).
JavaScript only has one Number type so is technically neither a float or an integer.
However this isn't really relevant as the value (or rather representation of it) is not specific to JavaScript and uses E-Notation which is a standard way to write very large/small numbers.
Taking this in to account 2.2932600144518896e+160 is equivalent to 2.2932600144518896 * Math.pow(10,160) and approximately 229 followed by 158 zeroes i.e. very flippin' big.

Math.sin() Different Precision between Node.js and C#

I have a problem in precision in the last digit after the comma.The javascript code generates one less Digit in compare with the C# code.
Here is the simple Node.js code
var seed = 45;
var x = Math.sin(seed) * 0.5;
console.log(x);//0.4254517622670592
Here is the simple C# code
public String pseudorandom()
{
int seed = 45;
double num = Math.Sin(seed) * (0.5);
return num.ToString("G15");//0.42545176226705922
}
How to achieve the same precision?
The JavaScript Number type is quite complex. It looks like floating point number will probably be like IEEE 754-2008 but some aspects are left to the implementation. See http://www.ecma-international.org/ecma-262/6.0/#sec-number-objects sec 12.7.
There is a note
The output of toFixed may be more precise than toString for some
values because toString only prints enough significant digits to
distinguish the number from adjacent number values. For example,
(1000000000000000128).toString() returns "1000000000000000100", while
(1000000000000000128).toFixed(0) returns "1000000000000000128".
Hence to get full digit accuracy you need something like
seed = 45;
x = Math.sin(seed) * 0.5;
x.toFixed(17);
// on my platform its "0.42545176226705922"
Also, note the specification for how the implementation of sin and cos allow for some variety in the actual algorithm. It's only guaranteed to within +/- 1 ULP.
Using java the printing algorithm is different. Even forcing 17 digits gives the result as 0.42545176226705920.
You can check you are getting the same bit patterns using x.toString(2) and Double.doubleToLongBits(x) in Java.
return num.ToString("G15");//0.42545176226705922
actually returns "0.425451762267059" (no significant digit + 15 decimal places in this example), and not the precision shown in the comment after.
So you would use:
return num.ToString("G16");
to get "0.4254517622670592"
(for your example - where the significant digit is always 0) G16 will be 16 decimal places.

Math inaccuracy in Javascript: safe to use JS for important stuff?

I was bored, so I started fidlling around in the console, and stumbled onto this (ignore the syntax error):
Some variable "test" has a value, which I multiply by 10K, it suddenly changes into different number (you could call it a rounding error, but that depends on how much accuracy you need). I then multiply that number by 10, and it changes back/again.
That raises a few questions for me:
How in accurate is Javascript? Has this been determined? I.e. a number that can be taken into account?
Is there a way to fix this? I.e. to do math in Javascript with complete accuracy (within the limitations of its datatype).
Should the changed number after the second operation be interpreted as 'changing back to the original number' or 'changing again, because of the inaccuracy'?
I'm not sure whether this should be a separate question, but I was actually trying to round numbers to a certain amount after the decimal point. I've researched it a bit, and have found two methods:
> Method A
function roundNumber(number, digits) {
var multiple = Math.pow(10, digits);
return Math.floor(number * multiple) / multiple;
}
> Method B
function roundNumber(number, digits) {
return Number(number.toFixed(digits));
}
Intuitively I like method B more (looks more efficient), but I don't know what going on behind the scenes so I can't really judge. Anyone have an idea on that? Or a way to benchmark this? And why is there no native round_to_this_many_decimals function? (one that returns an integer, not a string)
How in accurate is Javascript?
Javascript uses standard double precision floating point numbers, so the precision limitations are the same as for any other language that uses them, which is most languages. It's the native format used by the processor to handle floating point numbers.
Is there a way to fix this? I.e. to do math in Javascript with complete accuracy (within the limitations of its datatype).
No. The precision limitations lies in the way that the number is stored. Floating point numbers doesn't have complete accuracy, so no matter how you do the calculations you can't achieve absolute accuracy as the result goes back into a floating point number.
If you want complete accuracy then you need to use a different data type.
Should the changed number after the second operation be interpreted as
'changing back to the original number' or 'changing again, because of
the inaccuracy'?
It's changing again.
When a number is converted to text to be displayed, it's rounded to a certain number of digits. The numbers that look like they are exact aren't, it's just that the limitations in precision doesn't show up.
When the number "changes back" it's just because the rounding again hides the limitations in the precision. Each calculation adds or subtracts a small inaccuracy in the number, and sometimes it just happens to take the number closer to the number that you had originally. Eventhough it looks like it's more accurate, it's actually less accurate as each calculation adds a bit of uncertainty.
Internally, JavaScript uses 64-bit IEEE 754 floating-point numbers, which are a widely used standard and usually guarantee about 16 digits of accuracy. The error you witnessesed was on the 17th significant digit of the number and was reeeally tiny.
Is there a way to [...] do math in Javascript with complete accuracy (within the limitations of its datatype).
I would say that JavaScript's math is completely accurate within the limitations of its datatype. The error you witnessed was outside of those limitations.
Are you working with calculations that require a higher degree of precision than that?
Should the changed number after the second operation be interpreted as 'changing back to the original number' or 'changing again, because of the inaccuracy'?
The number never really became more or less accurate than the original value. It was only when the value was converted into a decimal value that a rounding error became apparent. But this was not a case of the value "changing back" to an accurate number. The rounding error was just too small to display.
And why is there no native round_to_this_many_decimals function? (one that returns an integer, not a string)
"Why is the language this way" questions are not considered very productive here, but it is easy to get around this limitation (assuming you mean numbers and not integers). This answer has 337 upvotes: +numb.toFixed(digits);, but note that if you try to display a number produced with that expression, there's no guarantee that it will actually display with only six digits. That's probably one of the reasons why JavaScript's "round to N places" function produces a string and not a number.
I came across the same few times and with further research I was able solve the little issues by using the library below
Math.js Library
Sample
import {
atan2, chain, derivative, e, evaluate, log, pi, pow, round, sqrt
} from 'mathjs'
// functions and constants
round(e, 3) // 2.718
atan2(3, -3) / pi // 0.75
log(10000, 10) // 4
sqrt(-4) // 2i
pow([[-1, 2], [3, 1]], 2) // [[7, 0], [0, 7]]
derivative('x^2 + x', 'x') // 2 * x + 1
// expressions
evaluate('12 / (2.3 + 0.7)') // 4
evaluate('12.7 cm to inch') // 5 inch
evaluate('sin(45 deg) ^ 2') // 0.5
evaluate('9 / 3 + 2i') // 3 + 2i
evaluate('det([-1, 2; 3, 1])') // -7
// chaining
chain(3)
.add(4)
.multiply(2)
.done() // 14

Math.pow() not giving precise answer

I've used Math.pow() to calculate the exponential value in my project.
Now, For specific values like Math.pow(3,40), it returns 12157665459056929000.
But when i tried the same value using a scientific Calculator, it returns 12157665459056928801.
Then i tried to traverse the loop till the exponential value :
function calculateExpo(base,power){
base = parseInt(base);
power = parseInt(power);
var output = 1;
gameObj.OutPutString = ''; //base + '^' + power + ' = ';
for(var i=0;i<power;i++){
output *= base;
gameObj.OutPutString += base + ' x ';
}
// to remove the last comma
gameObj.OutPutString = gameObj.OutPutString.substring(0,gameObj.OutPutString.lastIndexOf('x'));
gameObj.OutPutString += ' = ' + output;
return output;
}
This also returns 12157665459056929000.
Is there any restriction to Int type in JS ?
This behavior is highly dependent on the platform you are running this code at. Interestingly even the browser matters even on the same very machine.
<script>
document.write(Math.pow(3,40));
</script>
On my 64-bit machine Here are the results:
IE11: 12157665459056928000
FF25: 12157665459056929000
CH31: 12157665459056929000
SAFARI: 12157665459056929000
52 bits of JavaScript's 64-bit double-precision number values are used to store the "fraction" part of a number (the main part of the calculations performed), while 11 bits are used to store the "exponent" (basically, the position of the decimal point), and the 64th bit is used for the sign. (Update: see this illustration: http://en.wikipedia.org/wiki/File:IEEE_754_Double_Floating_Point_Format.svg)
There are slightly more than 63 bits worth of significant figures in the base-two expansion of 3^40 (63.3985... in a continuous sense, and 64 in a discrete sense), so hence it cannot be accurately computed using Math.pow(3, 40) in JavaScript. Only numbers with 52 or fewer significant figures in their base-two expansion (and a similar restriction on their order of magnitude fitting within 11 bits) have a chance to be represented accurately by a double-precision floating point value.
Take note that how large the number is does not matter as much as how many significant figures are used to represent it in base two. There are many numbers as large or larger than 3^40 which can be represented accurately by JavaScript's 64-bit double-precision number values.
Note:
3^40 = 1010100010111000101101000101001000101001000111111110100000100001 (base two)
(The length of the largest substring beginning and ending with a 1 is the number of base-two significant figures, which in this case is the entire string of 64 digits.)
Haskell (ghci) gives
Prelude> 3^40
12157665459056928801
Erlang gives
1> io:format("~f~n", [math:pow(3,40)]).
12157665459056929000.000000
2> io:format("~p~n", [crypto:mod_exp(3,40,trunc(math:pow(10,21)))]).
12157665459056928801
JavaScript
> Math.pow(3,40)
12157665459056929000
You get 12157665459056929000 because it uses IEEE floating point for computation. You get 12157665459056928801 because it uses arbitrary precision (bignum) for computation.
JavaScript can only represent distinct integers to 253 (or ~16 significant digits). This is because all JavaScript numbers have an internal representation of IEEE-754 base-2 doubles.
As a consequence, the result from Math.pow (even if was accurate internally) is brutally "rounded" such that the result is still a JavaScript integer (as it is defined to return an integer per the specification) - and the resulting number is thus not the correct value, but the closest integer approximation of it JavaScript can handle.
I have put underscores above the digits that don't [entirely] make the "significant digit" cutoff so it can be see how this would affect the results.
................____
12157665459056928801 - correct value
12157665459056929000 - closest JavaScript integer
Another way to see this is to run the following (which results in true):
12157665459056928801 == 12157665459056929000
From the The Number Type section in the specification:
Note that all the positive and negative integers whose magnitude is no greater than 253 are representable in the Number type ..
.. but not all integers with large magnitudes are representable.
The only way to handle this situation in JavaScript (such that information is not lost) is to use an external number encoding and pow function. There are a few different options mentioned in https://stackoverflow.com/questions/287744/good-open-source-javascript-math-library-for-floating-point-operations and Is there a decimal math library for JavaScript?
For instance, with big.js, the code might look like this fiddle:
var z = new Big(3)
var r = z.pow(40)
var str = r.toString()
// str === "12157665459056928801"
Can't say I know for sure, but this does look like a range problem.
I believe it is common for mathematics libraries to implement exponentiation using logarithms. This requires that both values are turned into floats and thus the result is also technically a float. This is most telling when I ask MySQL to do the same calculation:
> select pow(3, 40);
+-----------------------+
| pow(3, 40) |
+-----------------------+
| 1.2157665459056929e19 |
+-----------------------+
It might be a courtesy that you are actually getting back a large integer.

Categories