Monday, December 12, 2022

Multiplicative square roots

I will just discuss briefly the first section of a paper by William Waterhouse (2012), “Square Root as a Homomorphism” (American Mathematical Monthly 119 (3), 235-239), which addresses the following question: given a field FF, when is it possible to define square roots for all squares compatibly with products, ie, so that ab=ab\sqrt {ab}=\sqrt a\,\sqrt b if a,bFa,b\in F are squares.

Real numbers. — Such a square root operation exists when FF is the field of real numbers: we are familiar with the process of taking the positive square root of a positive real number.

Finite fields. — It also exists in some finite fields. So let FF be a finite field, let qq be its number of elements; then qq is a power of a prime number pp, but if you wish, you may already assume that q=pq=p is prime. For simplicity, we assume that qq is odd. By Fermat's little theorem, every nonzero element aFa\in F satisfies aq1=1a^{q-1}=1. Then q1q-1 is even, we can write aq1=a(q1)/2)2=1a^{q-1}=a^{(q-1)/2})^2=1, so that a(q1)/2=±1a^{(q-1)/2}=\pm1, and Euler's criterion asserts that aa is a square if and only if a(q1)/2=1a^{(q-1)/2}=1. (That this condition is necessary is obvious: write a=b2a=b^2, one gets a(q1)/2=bq1=1a^{(q-1)/2}=b^{q-1}=1 by Fermat's criterion. Then, a counting argument shows that it is sufficient: the map bb2b\mapsto b^2 is 22 to 11 on nonzero elements, hence its image consists of (q1)/2(q-1)/2 elements, all of which are squares; since the polynomial equation T(q1)/2=1T^{(q-1)/2}=1 has at most (q1)/2(q-1)/2 solutions in FF, we obtained all of them in this way.)

For example, 1-1 is a square if and only if (1)(q1)/2=1(-1)^{(q-1)/2}=1, which happens if and only if (q1)/2(q-1)/2 is even, that is, q1(mod4)q\equiv 1\pmod 4. In this case, do we have a formula for a square root of 1-1? When q=pq=p, yes, but it is not an easy one: Wilson's theorem states that (p1)!1(modp)(p-1)!\equiv -1\pmod p, just because you may pair each integer aa such that 1<a<p11\lt a\lt p-1 with its multiplicative inverse modulo pp; then only two factors remain in the product and (p1)!1(p1)1(modp)(p-1)!\equiv 1\cdot (p-1)\equiv -1\pmod p. Now, we pair each integer aa such that 1ap11\leq a\leq p-1 with its additive inverse pap-a; we get (((p1)/2)!)2(1)((p1)/2)(((p-1)/2) ! )^2 (-1)^((p-1)/2), hence ((p1)/2)!)21(modp)((p-1)/2)!)^2\equiv -1\pmod p. This is not an easy formula, because computing the factorial takes a long time for large pp.

It is possible to do much quicker, but you need to have a die at your disposal. Indeed, choose an element aa such that 1ap11\leq a\leq p-1 and compute b=a(p1)/4b=a^{(p-1)/4}. Since b2=a(p1)/2=±1b^2=a^{(p-1)/2}=\pm1, two possibilities arise: when aa is a square, we get 11, but if aa is not a square, then we get 1-1. And if we choose aa randomly, we have one chance over two of not having chosen a square, hence one chance over two to get an element bb such that b2=1b^2=-1.

At this point you may ask why it isn't as long to compute the power a(p1)/4a^{(p-1)/4} than the factorial ((p1)/2)!((p-1)/2)!, and you would be right. The reason is that there is a fast recursive way to compute a power ana^n, by writing an=(a2)n/2a^n=(a^2)^{n/2} if nn is odd, and an=a(a2)(n1)/2a^n=a\cdot (a^2)^{(n-1)/2} if nn is odd. This leads to basically log2(n)\log_2(n) multiplications and squarings, and not nn multiplications (n1n-1, actually) as the naïve expression aaaa\cdot a\dots a might have let you think.

But let us go back to the question of computing square roots. As the last three paragraphs indicate, it could be difficult to do so when q1(mod4)q\equiv 1\pmod 4. However, it is extremly easy in the other case q3(mod4)q\equiv 3\pmod 4. Take a nonzero element aa which is a square, and write a(q1)/2=1a^{(q-1)/2}=1. Since q3(mod4)q\equiv 3\pmod 4, we write q=1+4mq=-1+4m so that a2m1=1a^{2m-1}=1, hence say a=a2m=(am)2a=a^{2m}=(a^m)^2. We have our square root, it is simply given by b=am=a(q+1)/4b=a^m=a^{(q+1)/4}. The resulting map, aama\mapsto a^m, gives us our desired multiplicative square roots on squares.

Complex numbers. — Now for a negative result, there is no multiplicative square root on the complex numbers, basically for the reason we have been taught that it leads to fallacies. All complex numbers are squares, so let us assume that we have a multiplicative square root r ⁣:CCr\colon \mathbf C\to\mathbf C. Letting i=r(1)i=r(-1), the contradiction comes from the relation i=r(i)2=r((i)2)=r(1)=i.-i = r(-i)^2=r((-i)^2)=r(-1)=i.

We can now state and prove Waterhouse's theorem:

Theorem.Let FF be a field (of characteristic 2\neq 2) and let SFS\subseteq F be the multiplicative monoid of squares. There exists a multiplicative homomorphism r ⁣:SFr\colon S\to F if and only if 1S-1\notin S.

Proof. — The same negative argument as in the complex numbers works whenever 1-1 is a square in FF. So let us assume that 1-1 is not a square and let us explain why a multiplicative square root exists. The proof, however, is not explicit but relies on some maximal principle. Moreover, we won't define the square root map directly, but its image.
Let us first analyse the situation. Assume that r ⁣:SFr\colon S\to F is a multiplicative square root. It is simpler to remove 00 from the discussion so we consider its restriction S×F×S^\times \to F^\times and still denote it by rr. By assumption, it is a morphism of groups, so that its image R×R^\times is a subgroup of F×F^\times. Observe that it does not contain 1-1, for if r(a)=1r(a)=-1, then a=r(a)2=(1)2=1a=r(a)^2=(-1)^2=1 but r(1)=1r(1)=1. Moreover, for every element aF×a\in F^\times, we have r(a2)2=a2r(a^2)^2=a^2, hence r(a2)=±ar(a^2)=\pm a, so that either aa, or a-a belongs to RR, but not both since 1∉R×-1\not\in R^\times. As a consequence, R×R^\times is a maximal subgroup of F×F^\times among those which do not contain 1-1: adding to R×R^\times any element aF×a\in F^\times such that aR×a\notin R^\times would lead to a subgroup R×,a\langle R^\times,a\rangle which contains 1-1.

Let us consider a maximal subgroup of F×F^\times containing the squares which does not contain 1-1. Starting from S×S^\times, which does not contain 1-1, this can be done using Zorn's lemma, or by transfinite induction: well ordering the elements of F×F^\times, and constructing R×R^\times by induction. Since R×R^\times contains the squares, the union R×aR×R^\times \cup a R^\times is a subgroup of F×F^\times; if it does not contain 1-1, then we replace R×R^\times by it, other wise we discard aa and keep R×R^\times.

Let aF×a\in F^\times. If aR×a\notin R^\times, the construction means that 1aR×-1\in aR^\times, hence aR×-a\in R^\times. But we can't have both aa and a-a in R×R^\times, for that would imply that 1R×-1\in R^\times.

If aF×a\in F^\times is a nonzero square, it has two square roots, of the form ±b\pm b, and we define r(a)r(a) to be its square root which belongs to R×R^\times. One has r(1)=1r(1)=1, because 1S×R×1\in S^\times\subset R^\times. For nonzero squares a,ba,b, the product r(a)r(b)r(a)r(b) is a square root of abab, and it belongs to R×R^\times, hence it equals r(ab)r(ab). This proves that the map rr is multiplicative. This concludes the proof.

Remark. — If you've studied some abstract algebra, you may have recognized something in the middle of the proof. Indeed, the quotient group V=F×/S×V=F^\times/S^\times has exponent 2: for every α\alpha in this group, α2=1\alpha^2=1. Consequently, even if it is written multiplicatively, this abelian group is a vector space over the field with 2-elements. Since 1-1 is not a square in F×F^\times, its class [1][-1] is nonzero in F×/S×F^\times/S^\times, and the quotient group W=R×/S×W=R^\times/S^\times is just a maximal vector subspace that does not contain [1][-1]. It is a hyperplane and is defined by a linear form ϕ\phi on VV. Since VV is written multiplicatively, this linear form corresponds to a group homomorphism f ⁣:F×{±1}f\colon F^\times \to\{\pm1\} which maps S×S^\times to 11 and such that f(1)=1f(-1)=-1. For every square a=b2a=b^2, we then have r(a)=bf(b)r(a)=b f(b).

In his paper, Waterhouse goes on by viewing “fields FF with a multiplicative square root rr” as a basic algebraic object, and considering such structures (F,r)(F,r) which can't be extended by adding algebraic elements. The final theorem of the paper shows that the Galois group Gal(F/F)\mathop{\rm Gal}(\overline F/F) is either cyclic of order 2, or is the additive group of the 2-adic integers.