5

I'm working through the exercises in Introduction to Boolean Algebra by Halmos and Givant. Looking to show the following, an exercise from the first chapter: every finite Boolean rng must have a unit.

Halmos and Givant attribute this observation to Stone, in his original paper on the representation theorem. Stone shows the result by explicitly constructing the unit as a sum of elementary symmetric polynomials. For a finite Boolean ring $B$ with $|B| = n$, we consider the elementary symmetric polynomials in $n$ variables, which are:

$e_1(x_1, \dots, x_n) = x_1 + \dots + x_n$,

$e_2(x_1, \dots, x_n) = x_1 x_2 + \dots + x_{n-1}x_n$, ...

$e_n(x_1, \dots, x_n) = x_1\cdots x_n$

Stone shows that the unit in a finite Boolean rng (with $n$ elements) is given as the sum of $e_1(a_1, \dots, a_n) + \dots + e_n(a_1, \dots, a_n)$ where $\{a_1, \dots, a_n\} = B$.

What I'm wondering, is how one could figure this out? What observations about Boolean rings do I need to make in order to know that the above construction is the right one to try?

As a side note, I see that in this post, Martin Brandenburg gives another argument for the conclusion using the Nakayama lemma (which seems a little high-powered for this exercise) and mentions that a direct argument is possible.

  • 5
    Wow, that's a weird construction. For what it's worth conclusion that a finite Boolean rng is a ring is totally obvious from the perspective of Boolean algebras as lattices (though of course it takes a bit of work to deduce the relevant order-theoretic properties from the definition of a Boolean rng). – Eric Wofsey Feb 22 '20 at 04:39

3 Answers3

5

Let me first give a much more intuitive argument for why a finite Boolean rng is a ring, and then relate it to Stone's construction. Any Boolean rng is in particular a lattice (bounded below but not necessarily above), with the lattice operations defined by $$x\wedge y=xy$$ and $$x\vee y=x+y+xy.$$ Since multiplication is the same as the lattice join operation, a multiplicative identity is the same thing as a greatest element of the lattice. But now the conclusion is obvious: in a finite lattice, you can find a greatest element by just taking the join of all the elements.

Now where does Stone's construction come in? It turns out that if you translate an $n$-fold join back into the rng operations, you just get the sum of the elementary symmetric polynomials. This is clear for $n=2$ from the definition: $$x\vee y=(x+y)+xy=e_1(x,y)+e_2(x,y).$$ If you calculate out what you get for $n=3$, it is not hard to guess that it works for any $n$, and then it's also not hard to prove it by induction on $n$.

This is also closely related to inclusion-exclusion (which is another way you could come up with it): to count each element of an $n$-fold union $x_1\cup\dots \cup x_n$ exactly once, inclusion-exclusion tells you to first count each element of each of $x_1,x_2,\dots,x_n$ once, then subtract each element of each binary intersection, then add in each element of each ternary intersection, and so on. In other words, if you let $1_x$ denote the characteristic function of a set $x$, then the equation $$1_{x_1\cup\dots \cup x_n}=e_1(1_{x_1},\dots,1_{x_n})-e_2(1_{x_1},\dots,1_{x_n})+\dots+(-1)^{n-1}e_n(1_{x_1},\dots,1_{x_n})$$ is true (say, in the ring of all integer-valued functions on your universal set). Taking this mod $2$ so you are working with $\mathbb{F}_2$-valued functions which you can identify with the power set Boolean ring, this says exactly that $$x_1\vee\dots\vee x_n=e_1(x_1,\dots,x_n)+e_2(x_1,\dots,x_n)+\dots+e_n(x_1,\dots,x_n).$$ (Of course, this does not prove that this identity is true in an abstract Boolean rng unless you already have something like the representation theorem, but it's a way you might come up with it as a guess.)

Yet another way you can come up with this formula: the join operation is just what you get by conjugating the meet operation by the complement operation (of course this only makes sense when you have a unit). In terms of ring operations, the complement operation is $x\mapsto 1+x$, so this says that $x_1\vee \dots \vee x_n$ should be given by $$1+(1+x_1)(1+x_2)\dots(1+x_n).$$ When you expand out that product you'll just get the sum of all possible products of some of the $x_i$'s, and then when you add $1$ you'll remove the empty product so you're just left with the sum of the elementary symmetric polynomials. (If you use the same idea but with integer-valued functions where the complementation operation is $x\mapsto 1-x$, this gives a slick proof of inclusion-exclusion.)

Eric Wofsey
  • 342,377
  • Thanks! I imagine Stone was led to the construction by just considering $n$-fold joins, which isn't particularly satisfying, but so it goes. I suppose Halmos and Givant might have had some other proof (which doesn't go via symmetric polynomials) in mind, since this comes before they even mention the words "Boolean algebra." – omegaplusone Feb 22 '20 at 05:29
  • Well, it's pretty easy to prove directly that if you multiply $e_1+\dots+e_n$ by one of the $x_i$ you'll get $x_i$, so that it is a multiplicative identity (just think about what the terms are and you can figure out that all but the $x_i$ term cancel out in pairs). That is presumably the proof Halmos and Givant expect you to find. – Eric Wofsey Feb 22 '20 at 05:39
4

Inspired by Eric Wofsey's last comment, I think the best way to carry this argument out is to reason via the complementation operation in the unitization of a Boolean rng. A previous exercise was to show that this always exists.

So let $B$ be some finite non-degenerate Boolean ring and consider its unitization $\mathbb{Z}_2 \oplus B$. Then observe that for each $a_j \in B$, $a_j + a_j = a_j + a_j^2 = a_j(1 + a_j) = 0$. So consider the product $$ \prod_{i = 1}^n (1 + a_i) $$

Then for each $a_j$, we have $a_j(\prod_i (1 + a_i)) = 0$, and so we also have, for each $a_j \in B$ $$ a_j\left[1 + \prod_{i = 1}^n (1 + a_i)\right] = a_j $$ Observe that $\prod_{i} (1 + a_i) = 1 + p(a_1, \dots, a_n)$, where $p$ is some non-monic polynomial. Since $p$ is not monic, it follows that $p(a_1, \dots, a_n)$ is an element of $B$. But then $$ \left[1 + \prod_{i = 1}^n (1 + a_i)\right] = (1 + 0) + (1 + p(a_1, \dots, a_n)) = (0 + p(a_1, \dots, a_n)) $$ So $p(a_1, \dots, a_n)$ is the unit in $B$.

4

This is related to the lattice-theoretic approach outlined in other solutions, but I feel like it is more elementary and doesn't require complete translation to meets and joins.

First note that $eR$ is a ring with identity $e$ for any $e\in R$. (Of course, commutativity is in play in a boolean rng.)

The thing is that if $f\notin eR$, then $g=e+f+ef$ is an element such that $gR\supseteq (f, eR)$.

What has happened? Starting at any $eR$, we've shown that if $eR\neq R$, then $eR$ is contained in a larger ring $gR$ with identity $g$.

Of course, since $R$ is finite, this can't go on forever: at some point, $gR=R$, and $R$ has identity $g$.

rschwieb
  • 160,592