The Euclidean GCD extended binary argorithm (https://en.wikipedia.org/wiki/Extended_Euclidean_algorithm) is a mechanism for finding the lowest common factor of 2 numbers, and also finding two weights that, when multiplied by the 2 original numbers, produce the common factor.
It works because we pick transformations of the 2 numbers that preserve the lowest common factor. For instance in the traditional Euclidean algorithm, we reduce the value by performing:
Rn = Rn-2 - Q.Rn-1
where values R are the results of applying the GCD preserving transformation (R0 = a, R1 = b, the initial values).
Q is any value, but typically the Euclidean quotient (integer part) of Rn-2 / Rn-1. This is often written as
Rn-2 (mod Rn-1).
This can be seen to be true because if R0 = a = md, and R1 = b = nd, where d is the common factor and m and n
are co-prime then:
R2 = R0 - Q.R1
R2 = md - Q.nd
R2 = d(m - Q.n)
Therefore R2 is also a multiple of d. m and n must be co-prime because if they shared a common factor then
this could be extracted from both and multiplied into the GCF d.
By repeating this transformation, we can reduce the values of Rn, until one of them becomes 0, when we know that
the previous one was the GCF d.
The quotient Q is expensive for binary machines and instead a binary algorithm can be used. In the
binary algorithm, we first remove all common 2s from the prime factors. For instance if a = 2^4 . 3^2 . 5, and
b = 2^3 . 5 . 7 then we remove 2^3 from both numbers. Eventually we will multiply the 2s back into the GCD.
Now we have 2 numbers, a and b, one of which is definitely odd, because one of them will have no powers of 2 in
their prime factorization. We will call the odd one a, and if that isn't the case we will swap the labels so that
it is.
b is either even or odd. If it is even, then we can discard the factor of 2 because a does not have it in
common. If it is odd, then we can apply a transformation to either a or b that preserves the GCF, but produces
an even number.
Rn = Rn-1 - Rn-2
Since both Rn-1 (a) and Rn-2 (b) are odd, the result of the difference is even. We can then remove the factor
of 2 as we did above.
We can see that Rn = Rn-1 - Rn-2 preserves the GCF. If Rn-1 = md and Rn-2 = nd, where m and n are co-prime
as before, then:
Rn = md - nd
Rn = d . (m - n)
In additon to finding d, the GCF, we would like to find two integers s and t which satisfy:
d = s.a + t.b
where a and b are the initial values.
We can find these by starting with two sets of s and t (one for each starting value), and tracking them
through the transformations:
Rx = a.Sx + b.Tx
Ry = a.Sy + b.Ty
If Rx = a and Ry = b, then it should be obvious that:
Rx = a.1 + b.0. ie Sx=1, Tx=1
Ry = a.0 + b.1. ie Sx=0, Tx=1
When, in the binary algorithm, we subtract the 2 values, we have:
Rx -> Rx - Ry
which means that:
a.Sx + b.Tx -> a.Sx + b.Tx - a.Sy - b.Ty
a.Sx + b.Tx -> a.Sx - a.Sy + b.Tx - b.Ty
a.Sx + b.Tx -> a.(Sx - Sy) + b.(Tx - Ty)
therefore when Rx -> Rx - Ry
Sx -> Sx - Sy
Tx -> Tx - Ty
In the case where we discard the factor of two, it is slightly more complex.
Rx -> Rx / 2
a.Sx + b.Tx -> (a.Sx + b.Tx) / 2 -> a.(Sx/2) + b.(Tx/2)
Sx -> Sx/2
Tx -> Tx/2
This only works if Sx and Tx are both even (they must remain integers).
In the case where they are not both even, we can use a different transformation for Sx and Tx. We
know that Rx is even, and at most one of a and b are even, because we discarded common powers of 2 from a and b
before we started these transformations. So, if a is odd then Sx is even, and if b is odd then
Tx is even. If a and b are both odd, then so are Sx and Tx.
Acceptable combinations when Sx and Tx are not both even:
Rx(even) = a(odd) . Sx(odd) + b(odd) . Tx(odd)
Rx(even) = a(odd) . Sx(even) + b(even) . Tx(odd)
Rx(even) = a(even) . Sx(odd) + b(odd) . Tx(even)
So we use the following transformation when Sx and Tx are not both even.
Sx -> (Sx + b)/2
Tx -> (Tx - a)/2
You can see that Sx + b must be even, and Tx - a must also be even.
And you can see that the transformation produces the correct division by 2 by substituting the new Sx and Tx into the Rn transformation:
a.Sx + b.Tx -> a.(Sx + b)/2 + b.(Tx - a)/2
a.Sx + b.Tx -> a.Sx/2 + a.b/2 + b.Tx/2 - b.a/2
a.Sx + b.Tx -> a.Sx/2 + b.Tx/2 + (a.b/2 - a.b/2)
a.Sx + b.Tx -> a.Sx/2 + b.Tx/2
a.Sx + b.Tx -> (a.Sx + b.Tx) / 2
Rn -> Rn / 2
1) Start with 'a' and 'b'
2) Remove common factors of 2 from both and remember how many factors of 2 there were (=c)
3) Make 'Rx' = 'a'
4) Make `Ry` = 'b'
5) Set 'Sx' = '1', 'Sy' = '0', 'Tx' = '0', 'Ty' = '1'
6) Repeat until Rx==Ry:
7) If 'Rx' is odd:
8) If (Rx > Ry):
9) Rx -> Rx - Ry
10) Sx -> Sx - Sy
11) Tx -> Tx - Ty
12) Else:
13) Ry -> Ry - Rx
14) Sy -> Sy - Sx
15) Ty -> Ty - Tx
16) Else: (Rx is even)
17) Rx -> Rx / 2
18) If Sx==even AND Sy==even:
19) Sx -> Sx / 2
20) Tx -> Tx / 2
21) Else:
22) Sx -> (Sx + b) / 2
23) Tx -> (Tx - a) / 2
24) (at this point Rx == Ry == GCD without the common powers of 2)
25) GCD = Rx * 2^c
26) S = Sx
27) T = Tx
Performance is surprisingly good, in my opinion.
It produces the GCD and weights of two random 1024bit numbers in 2.4 seconds. Much of this seems to be JVM overhead as the GCD of 1 and 1 takes 2.28 seconds.