Saturday, March 29, 2025

A simple proof of a theorem of Kronecker

Kronecker's theorem of the title is the following.

Theorem.Let αC\alpha\in\mathbf C be an algebraic integer all of whose conjugates have absolute value at most 11. Then either α=0\alpha=0, or α\alpha is a root of unity.

This theorem has several elementary proofs. In this post, I explain the simple proof proposed by Gebhart Greiter in his American Mathematical Monthly note, adding details so that it would (hopefully) be more accessible to students.

The possibility α=0\alpha=0 is rather unentertaining, hence let us assume that α0\alpha\neq0.

Let us first analyse the hypothesis. The assumption that α\alpha is an algebraic integer means that there exists a monic polynomial fZ[T]f\in\mathbf Z[T] such that f(α)=0f(\alpha)=0. I claim that ff can be assumed to be irreducible in Q[T]\mathbf Q[T]; it is then the minimal polynomial of α\alpha. This follows from Gauss's theorem, and let me give a proof for that. Let gQ[T]g\in \mathbf Q[T] be a monic irreducible factor of ff and let hQ[T]h\in\mathbf Q[T] such that f=ghf=gh. Chasing denominators and dividing out by their gcd, there are polynomials g1,h1Z[T]g_1,h_1\in\mathbf Z[T] whose coefficients are mutually coprime, and natural integers u,vu,v such that g=ug1g=ug_1 and h=vh1h=vh_1. Then (uv)f=g1h1(uv)f=g_1 h_1. Since ff is monic, this implies that uvuv is an integer. Let us prove that uv=1uv=1; otherwise, it has a prime factor pp. Considering the relation (uv)f=g1h1(uv)f=g_1h_1 modulo pp gives 0=g1h1(modp)0=g_1 h_1 \pmod p. Since their coefficients are mutually coprime, the polynomials g1g_1 and h1h_1 are nonzero modulo pp, hence their product is nonzero. This is a contradiction.

So we have a monic polynomial fZ[T]f\in\mathbf Z[T], irreducible in Q[T]\mathbf Q[T], such that f(α)=0f(\alpha)=0. That is to say, ff is the minimal polynomial of α\alpha, so that the conjugates of α\alpha are the roots of ff. Note that the roots of ff are pairwise distinct — otherwise, the gcd(f,f)\gcd(f,f') would be a nontrivial factor of ff. Moreover, 00 is not a root of ff, for otherwhise one could factor f=Tf1f=T\cdot f_1.

Let us now consider the companion matrix to ff: writing f=Tn+c1Tn1++cnf=T^n+c_1T^{n-1}+\dots+c_n, so that n=deg(f)n=\deg(f), this is the matrix  Cf=(00cn1cn10cn201c1). C_f = \begin{pmatrix} 0 & \dots & 0 & -c_n \\ 1 & \ddots & \vdots & -c_{n-1} \\ 0 & \ddots & & -c_{n-2} \\ 0 & \cdots & 1 & -c_1 \end{pmatrix}. If e1,,ene_1,\ldots,e_n are the canonical column vectors e1=(1,0,,0)e_1=(1,0,\dots,0), etc., then Cfe1=e2C_f\cdot e_1=e_2, \ldots, Cfen1=enC_f \cdot e_{n-1}=e_{n}, and Cfen=cne1c1enC_f\cdot e_n = -c_{n} e_1-\dots -c_1 e_n. Consequently,  f(Cf)e1=Cfne1+c1Cfn1e1++cne1=0. f(C_f)\cdot e_1 = C_f^n\cdot e_1 +c_1 C_f^{n_1} \cdot e_1+ \dots +c_n e_1 = 0. Moreover, for 2kn2\leq k\leq n, one has f(Cf)ek=f(Cf)Cfk1e1=Cfk1f(Cf)e1=0f(C_f)\cdot e_k=f(C_f)\cdot C_f^{k-1}\cdot e_1=C_f^{k-1}\cdot f(C_f)\cdot e_1=0. Consequently, f(Cf)=0f(C_f)=0 and the complex eigenvalues of f(Cf)f(C_f) are roots of ff. Since ff has simple roots, CfC_f is diagonalizable and their exists a matrix PGL(n,C)P\in\mathrm{GL}(n,\mathbf C) and diagonal matrix DD such that Cf=PDP1C_f=P\cdot D\cdot P^{-1}, the diagonal entries of DD are roots of ff. Since 00 is not a root of ff, the matrix DD is invertible, and CfC_f is invertible as well. More precisely, one can infer from the definition of CfC_f that g(Cf)e10g(C_f)\cdot e_1\neq 0 for any nonzero polynomial gg of degre <n<n, so that ff is the minimal polynomial of CfC_f. Consequently, all of its roots are actually eigenvalues of CfC_f, and appear in DD; in particular, α\alpha is an eigenvalue of CfC_f.

For every k1k\geq 1, one has Cfk=P(Dk)P1C_f^k=P\cdot (D^k)\cdot P^{-1}. Since all entries of DD have absolute value at most 1,1, the set of all DkD^k is bounded in Mn(C)\mathrm{M}_n(\mathbf C). Consequently, the set {Cfk;kZ}\{C_f^k\,;\, k\in\mathbf Z\} is bounded in Mn(C)\mathrm M_n(\mathbf C). On the other hand, this set consists in matrices in Mn(Z)\mathrm M_n(\mathbf Z). It follows that this set is finite.

There are integers kk and \ell such that k<k< \ell and Cfk=CfC_f^k=C_f^\ell. Since CfC_f is invertible, one has Cfk=1C_f^{\ell-k}=1. Since α\alpha is an eigenvalue of Cf,C_f, this implies αk=1\alpha^{\ell-k}=1. We thus have proved that α\alpha is a root of unity.