Binary XGCD

[erik-3milabs]

Continuation of the work started in #755

[dvdplm]

Can you elaborate a bit on where this restriction comes from and if it's a hard requirement or not? Skimming the example code for the paper doesn't seem to mention this.

It's a pretty limiting restriction; I think many applications will want to use full-fat uints with the MSB set.

[erik-3milabs]

Can you elaborate a bit on where this restriction comes from and if it's a hard requirement or not? Skimming the example code for the paper doesn't seem to mention this.

It's a pretty limiting restriction; I think many applications will want to use full-fat uints with the MSB set.

I was encountering integer overflows whenever the top bit was set. However, in the past two days I learned a fact and a trick that we can use:

The trick

Currently, I'm using an Ints inside the matrix representing the xgcd state. However, I learned this interesting fact that the signs of this state matrix will always occur in one of the following two patterns:

[ + - ]       [ - + ]
[ - + ]   or  [ + - ]

Because of this, I can refactor this matrix to use Uints instead, as long as it also holds a pattern: ConstChoice. Then, the multiplication M3 = M1 x M2 is computed as a standard matrix multiplication and M3.pattern = M1.pattern.eq(M2.pattern).

This saves one bit in the representation and, as a result, should prevent overflows.

The fact.

This morning, I learned that if a does not divide b (or the other way around), then the Bézout coefficients x,y such that ax + by = gcd are such that $$|x| \leq \lfloor\frac{b}{2 * gcd}\rfloor$$ and $$|y| \leq \lfloor\frac{a}{2 * gcd}\rfloor$$.

These two things combined should give us ample bits to compute the Bézout coefficients without overflowing.

I hope to work on this soon!

[dvdplm]

That is good news! Thank you for explaining.

Can you point me to a source on “the fact”?

[erik-3milabs]

But of course! It is stated as fact on Extended GCD Wikipedia page, at the end of the Description section. Not the world's greatest source, but all my experiments point in the same direction 🙈