People expect computers to be entirely accurate and precise with numbers. You know, you grow up with calculators. You expect that if you type in: "0.
1 + 0. 2", you expect it to come out to "0. 3".
And sooner or later every programmer, as they're learning, tries something like that. And they type it into a programming language. Something like "0.
1 + 0. 2", and they will hit Enter, and what the computer will give them is something like "0. 300000000.
. . 1".
And at that point, they will have started to learn about floating point numbers, and they will be starting to tear their hair out. Floating point numbers are essentially scientific notation. For the uninitiated, let's say you're an astrophysicist, and you have the speed of light, which is this enormous number with eight 0s after it.
And you want to multiply it by a tiny distance. It doesn't really matter what we're measuring here, it's just a big number and a small number. Let's say, 0.
00000015. Let's say you want to multiply these two together. (I may have made this a bit deliberately easy.
) You could try and do some really complicated maths and work it all out, but you don't need to. Because in scientific notation, this is 3 times 10 to the 8, and this one is 1. 5 times 10 to the -7.
So you don't need to do anything particularly complicated to multiply these two together. All you have to do is multiply these, which is 4. 5, and then add these exponents together here.
So, 8 add -7, eight minus seven means that's 1. So we'll have 4. 5 times 10 to the 1, or 4.
5 times 10, or 45. Brilliant. That's much easier than trying to actually handle large numbers.
And it also brings us to the concept of significant figures. Because the speed of light isn't exactly 3 times 10 to the 8. It's 2.
997[something], and there's a lot of digits on the end of there which don't really matter for everyday calculations. We round it up. We say it's 3 times 10 to the 8.
Or maybe, for being precise, we say it's 2. 997 times 10 to the 8, but we don't really worry about the numbers beyond that. Hold that thought.
So, we'll come back to significant digits later. The two main advantages to floating point are speed and efficiency. Speed, because floating point has been built over many, many years.
It is now lightning fast for computers to deal with. And efficiency, because it can deal with really big numbers -- massive numbers, "size of the universe" numbers -- and really small numbers -- "size of an atom" numbers -- without needing enormous amounts of space. If you had a format that needed to include the size of the universe and the size of an atom, you'd need so many zeroes stored on both ends or you'd need something particularly complicated for a computer to work out that it would either be really inefficient, storing one number as a megabyte, or you would need something that would be really fiddly for a computer to work out as it tries to place all the zeroes and the numbers everywhere.
Let's talk about base 10. Let's talk about us - people. (Let's talk about the lights being on all of a sudden.
That's weird. ) In base 10, in our normal human numbers, so you know ten digits - base 10, we have hundreds, tens, and units. You learn that in primary school.
And over here, you have tenths, hundredths, and thousandths. So if you want to say "one tenth" in decimal, you say "0. 1", and that's normal, and that's natural.
Base 2, on the other hand, binary. . .
, computers. . .
, they don't do that. Because what they see are fours, twos, and ones. And then over here: halves, quarters, eighths and sixteenths.
There's no 0. 1 in there. In fact, 0.
1 in binary is "0. 00011", and that "0011" is recurring. It's actually 0.
00011001100110011. . .
all the way to infinity. Which is interesting, because 32-bit computers, the ones we generally use, (we've moved a bit on now) 32-bit computers only store 23 significant digits. They also store where the decimal point is.
They basically do scientific notation in base 2. That's what floating point is. So they say, "All right, so we've got this long binary number, times 2 to the something.
" And that's a problem. Because what it loses is precision. It doesn't understand recursion.
I have an analogy for that in base 10: if you try to write "a third" as a decimal, we go: hundreds, tens, units, tenths, hundredths, thousandths, and then, we try and do this and we end up with 0. 333333333. .
. recurring. Now imagine you don't understand recurring numbers, like a computer, like scientific notation if you take away that recursion, like floating point arithmetic doesn't.
So this is where we bring everything together. If you're in base 10, if you're in decimal, let's say you have 1/3 + 1/3 + 1/3. As humans we go, "Well that's 0.
3 recurring, plus 0. 3 recurring, plus 0. 3 recurring, which is 1.
" But say you're a computer and you don't understand recursion, because floating point math is essentially significant figures. So a computer would look at that and go, "0. 333333333.
. . plus 333.
. . , plus 333.
. . ", and it will say, "Well, a third plus a third plus a third.
Total it to 0. 99999999. .
. " But after a while, you'll stop. Because you've run out of digits.
What you've just got there is a floating point rounding error. Let's take it down again to base 2. Let's do 1/10 + 2/10.
Which is 0. 00011001100110011. .
. Oh! We've run out of digits!
Because floating point arithmetic on 32-bit computers only stores 23 digits. And then you try and add 2/10, which is 0. 001100110011.
. . I'm not going to try and do that maths, because it's going to take ages but what happens is exactly the same as back here.
After 23 digits in 32-bit arithmetic (I think we've got 56 on 64-bit computers) it cuts. It doesn't understand recurring, which means that floating point looks at 1/10 + 2/10, what you just typed in, "0. 1 + 0.
2" in decimal, and it says that 1/10 + 2/10 does NOT quite equal 3/10. Because, to its mind, it doesn't. For nearly all cases, it IS close enough.
In most cases, you don't need more than 23 binary digits of precision. If you do, you go to 64-bit. If you are doing a 3D game and you need to know where to point something on the screen, it doesn't matter if it's 1/100000 of a pixel out because it will be in the right place.
And if you're doing big calculations, it doesn't matter if a tiny, tiny, tiny fraction -- something smaller than a molecule -- is out, because it will look more or less right: right beyond any precision or tolerance that the real world could possibly manage. But it means that if you're dealing with currency, and you say that you are adding 10p (or £0. 1) to 20p, every programmer hits this sooner or later, we just go like that.
That's easy. A computer will come back and say, "That's actually. .
. that many pence. " And in these particular human things, we will notice the error because suddenly to us it's very, very obvious.
That's the problem with floating point. To fix that, by the way, if you're doing currency, you either use a decimal type in more advanced programming languages or you just store everything as an integer in terms of pence, cents, or whatever and then divide by 100 at the end. Floating point is something that really, really confuses a lot of people.
It's tough to get your head around until you realize that it is just scientific notation. It's the same thing that we learned in maths class at school, only in base 2. It's halves and quarters and eighths, and once you crack that, once you realize that it's storing a set length of number and then working out where the decimal point could be and it's just significant figures, it starts to become a lot easier.
And you start to realize, "Oh, that's why if i type 0. 1 + 0. 2, what I get isn't 0.
3.
