Why is rounding necessary?

What happens when we write the following Julia code?

julia> x = 0.1
0.1

This appears to store the value 0.1 in a variable x of type Float64. In fact, however, it stores a slightly different number, since 0.1 cannot be represented exactly in binary floating point arithmetic, at any precision.

The true value that is actually stored in the variable can be conveniently determined in Julia using arbitrary-precision arithmetic with BigFloats:

julia> big(0.1)
1.000000000000000055511151231257827021181583404541015625000000000000000000000000e-01

So, in fact, the Julia float 0.1 refers to a real number that is slightly greater than 0.1. By default, such calculations are done in round-to-nearest mode (RoundNearest); i.e., the nearest representable floating-point number to 0.1 is used.

[Recall that to get a BigFloat that is as close as possible to the true 0.1, you can use a special string macro:

julia> big"0.1"
1.000000000000000000000000000000000000000000000000000000000000000000000000000002e-01

]

Suppose that we create a thin interval, containing just the floating-point number 0.1:

julia> II = Interval(0.1)
[0.1, 0.100001]

julia> showall(II)
Interval(0.1, 0.1)

It looks like II contains (the true) 0.1, but from the above discussion we see that it does not. In order to contain 0.1, the end-points of the interval must be rounded outwards ("directed rounding"): the lower bound is rounded down, and the upper bound is rounded up.

This rounding is handled by the @interval macro, which generates correctly-rounded intervals:

julia> a = @interval(0.1)
[0.0999999, 0.100001]

The true 0.1 is now correctly contained in the intervals, so that any calculations on these intervals will contain the true result of calculating with 0.1. For example, if we define

julia> f(x) = 2x + 0.2

then we can apply the function f to the interval a to obtain

julia> f(a)
[0.399999, 0.400001]

julia> showall(f(a))
Interval(0.39999999999999997, 0.4)

The result correctly contains the true 0.4.

More detail: the internal representation

Let's look at the internal representation of the Float64 number 0.1:

julia> bits(0.1)
"0011111110111001100110011001100110011001100110011001100110011010"

The last 53 bits of these 64 bits correspond to the binary expansion of 0.1, which is

0.000110011001100110011001100110011001100...

We see that the expansion is periodic; in fact, the binary expansion of 0.1 has an infinite repetition of the sequence of digits 1100. It is thus impossible to represent the decimal 0.1 in binary, with any precision.

The true value must be approximated by a floating-point number with fixed precision -- this procedure is called rounding. For positive numbers, rounding down may be accomplished simply by truncating the expansion; rounding up is accomplished by incrementing the final binary digit and propagating any resulting changes.