The Minimal Polynomial over $\mathbb{R}$ and the Real Jordan Form

Introduction

The minimal polynomial is the monic generator of the annihilator of $A$ in $\mathbb{R}[t]$, it divides $\chi_A$, and over $\mathbb{R}$ its irreducible factors are of degree one or two. The last clause is the whole difference from the complex case: a factor of degree two is a rotation block, and it cannot be split without leaving the field.

The article defines the minimal polynomial, proves that its roots are the spectrum and its exponents the sizes of the largest blocks, lists the irreducible factors over $\mathbb{R}$, states the real Jordan form with its rotation blocks, proves uniqueness, computes the dimension of the centraliser, and treats the companion matrix. It assumes The Element: Rank, Trace, Determinant, Characteristic Polynomial over $\mathbb{R}$ and The Elementary Symmetric Functions and the Real Characteristic Polynomial. It defers the rotation blocks as geometry to The Real Jordan Form and the Rotation Blocks and the spectral theorem to Normal Matrices over $\mathbb{R}$ and the Orthogonal Spectral Theorem.

The Annihilator and the Minimal Polynomial

The annihilator is a nonzero ideal of $\mathbb{R}[t]$, hence principal. Let

$$ \operatorname{Ann}(A)=\{p\in\mathbb{R}[t]:p(A)=0\}. $$

It is an ideal, and it is nonzero because $\chi_A\in\operatorname{Ann}(A)$ by Cayley–Hamilton. Since $\mathbb{R}[t]$ is a principal ideal domain, $\operatorname{Ann}(A)=(m_A)$ for a unique monic $m_A$, the minimal polynomial.

The minimal polynomial divides every annihilating polynomial, and in particular $\chi_A$. This is the definition of the generator. Hence $\deg m_A\le\deg\chi_A=n$, and $m_A=\chi_A$ exactly when no lower-degree polynomial annihilates $A$.

The minimal polynomial is a similarity invariant and is unchanged by transposition. $m_{SAS^{-1}}=m_A$ because $p(SAS^{-1})=Sp(A)S^{-1}$, and $m_{A^{\mathsf T}}=m_A$ because transposition is a ring anti-automorphism fixing $\mathbb{R}[t]$. So the minimal polynomial is a coarse invariant of the conjugacy class, strictly coarser than the characteristic polynomial only when the latter has repeated roots.

The algebra generated by $A$ has dimension $\deg m_A$. The powers $\mathrm{I},A,\dots,A^{d-1}$ with $d=\deg m_A$ are linearly independent, because a relation between them is an annihilating polynomial of degree less than $d$. They span $\mathbb{R}[A]$, because $m_A$ expresses $A^{d}$ in terms of the lower powers and induction handles the rest. So

$$ \dim_{\mathbb{R}}\mathbb{R}[A]=\deg m_A\le n . $$

Roots and Exponents

The roots of $m_A$ are exactly the eigenvalues of $A$. If $\lambda$ is an eigenvalue with eigenvector $v$ over $\mathbb{C}$, then $m_A(A)v=m_A(\lambda)v=0$, so $m_A(\lambda)=0$. Conversely a root $\mu$ of $m_A$ is a root of $\chi_A$ because $m_A$ divides $\chi_A$, hence an eigenvalue. So $m_A$ and $\chi_A$ have the same set of roots, and over $\mathbb{R}$ the same irreducible factors.

The exponent of a root is the size of the largest block for that root. Write the complex Jordan form with blocks of sizes $d_1\ge d_2\ge\dots$ for the eigenvalue $\lambda$. Then $(A-\lambda\mathrm{I})^{d_1}$ kills the block of size $d_1$ and all smaller ones, and no smaller power does. So the exponent of $t-\lambda$ in $m_A$ is $d_1$, while its exponent in $\chi_A$ is the sum of the sizes. At $n=3$ the nilpotent of type $(2,1)$ has $m_A=t^{2}$ and $\chi_A=t^{3}$, and the regular nilpotent has $m_A=\chi_A=t^{3}$.

The matrix is diagonalisable exactly when $m_A$ has simple roots. If $m_A$ is a product of distinct linear factors then $A$ is diagonalisable over $\mathbb{C}$; if in addition the factors are in $\mathbb{R}[t]$ and of degree one, then $A$ is diagonalisable over $\mathbb{R}$. If $m_A$ has a repeated factor, the Jordan form has a block of size at least two.

Over $\mathbb{R}$ the minimal polynomial need not split. For the rotation $J$ at $n=2$, $m_J=\chi_J=t^{2}+1$, irreducible over $\mathbb{R}$ and with simple roots $\pm i$. So $J$ is diagonalisable over $\mathbb{C}$ and not over $\mathbb{R}$, and the two statements are told apart by whether the factors of $m_A$ have degree one.

The Irreducible Factors over $\mathbb{R}$

Every monic irreducible polynomial over $\mathbb{R}$ has degree one or two. A polynomial of odd degree has a real root, so the irreducible factors are the linear $t-\lambda$ and the quadratics with negative discriminant. A monic irreducible quadratic is

$$ (t-a)^{2}+b^{2}=t^{2}-2at+(a^{2}+b^{2}),\qquad b\ne0, $$

with roots $a\pm bi$.

The minimal polynomial therefore factors as

$$ m_A(t)=\prod_{j}(t-\lambda_j)^{e_j}\prod_{k}\bigl((t-a_k)^{2}+b_k^{2}\bigr)^{f_k}, $$

with $\lambda_j\in\mathbb{R}$, $b_k\ne0$, and distinct factors. The first product collects the real eigenvalues, the second the conjugate pairs.

The exponents are the sizes of the largest blocks of each kind. For a real eigenvalue, $e_j$ is the size of the largest nilpotent block on the generalised eigenspace. For a conjugate pair, $f_k$ is the size of the largest block in the complex Jordan form for either member of the pair, and the two members give the same number because conjugation exchanges them.

The degree of $m_A$ counts the parameters of the algebra generated by $A$. A matrix with a real spectrum and distinct eigenvalues has $m_A=\chi_A$ of degree $n$; a rotation at $n=2$ has degree $2$ and generates a copy of $\mathbb{C}$ inside $M_2(\mathbb{R})$, which is the embedding of The Matrix Algebra $M_n(\mathbb{R})$.

The Real Jordan Form

The real Jordan form replaces each complex block by a real one of twice the size. For a real eigenvalue $\lambda$ and a block size $d$,

$$ J_d(\lambda)=\begin{pmatrix}\lambda&1&&\\&\ddots&\ddots&\\&&\ddots&1\\&&&\lambda\end{pmatrix}, $$

which is the ordinary Jordan block. For a conjugate pair $a\pm bi$ and a block size $d$,

$$ R_d(a,b)=\begin{pmatrix}C&I&&\\&\ddots&\ddots&\\&&\ddots&I\\&&&C\end{pmatrix},\qquad C=\begin{pmatrix}a&-b\\ b&a\end{pmatrix}, $$

a block of size $2d$, built from the rotation $C$ and the identity in the superdiagonal position.

Theorem. Every real matrix is similar over $\mathbb{R}$ to a direct sum of blocks $J_d(\lambda)$ and $R_d(a,b)$, and the list of blocks is unique up to order.

Proof. Regard $A$ as a complex matrix and take its complex Jordan form. The conjugation of $\mathbb{C}^{n}$ sends the generalised eigenspace of $\mu$ to that of $\bar\mu$ and conjugates the block structure, so the blocks for $\mu$ and for $\bar\mu$ have the same sizes. A pair of conjugate blocks of size $d$ is the complex form of one real block $R_d(a,b)$: on the real span of a Jordan chain and its conjugate, $A$ acts by $C$ on the pairs and by $I$ on the successive vectors of the chain. Real blocks for real eigenvalues are unchanged. Uniqueness follows from the uniqueness of the complex form together with the pairing.

The real form is a complex form with the conjugation made visible. The number of parameters is the same, the block sizes are the same, and only the presentation changes: the pair $\mu,\bar\mu$ becomes the pair $(a,b)$ of real numbers, and a complex block of size $d$ becomes a real block of size $2d$.

At $n=2$ the list is exhaustive and short. Either two real eigenvalues, giving $\operatorname{diag}(\lambda_1,\lambda_2)$ or a single block $J_2(\lambda)$, or a conjugate pair, giving $R_1(a,b)=C$ with $b\ne0$. The three kinds are the three orbits of the conjugation action up to the eigenvalues.

Uniqueness and the Centraliser

The centraliser of a matrix is computed block by block, and its dimension is read from the block sizes. For a single eigenvalue with block sizes $d_1,\dots,d_s$ the centraliser of the block has dimension

$$ \sum_{i,j=1}^{s}\min(d_i,d_j)=\sum_{r\ge1}(\lambda'_r)^{2}, $$

where $\lambda'$ is the conjugate partition of the partition given by the $d_i$, that is $\lambda'_r$ is the number of blocks of size at least $r$. The equality of the two expressions is a count of the pairs $(i,j)$ with $\min(d_i,d_j)\ge r$. For several distinct eigenvalues the dimensions add, because a matrix commuting with $A$ preserves each generalised eigenspace.

At $n=3$ the count separates the three nilpotents. For the regular nilpotent, one block of size $3$, the conjugate partition is $(1,1,1)$ and the dimension is $3$. For the type $(2,1)$ the conjugate partition is $(2,1)$ and the dimension is $5$. For $N=0$, three blocks of size $1$, the conjugate partition is $(3)$ and the dimension is $9$, the whole algebra. The three numbers $3,5,9$ are distinct, so the centraliser already separates the three classes.

At $n=2$ the two nilpotents are separated by the same count. The regular nilpotent has one block of size $2$, conjugate partition $(1,1)$, dimension $2$; the zero matrix has dimension $4$. A matrix with two distinct real eigenvalues has two blocks of size $1$, conjugate partition $(2)$, dimension $4$, and it is diagonalisable and nonderogatory at once.

The centraliser of a nonderogatory matrix is the algebra it generates. If $m_A=\chi_A$ then $\dim\mathbb{R}[A]=n=\dim Z(A)$, and since $\mathbb{R}[A]\subseteq Z(A)$ the two coincide. So for a matrix with a cyclic vector the commutant is the polynomial algebra, and the matrix is called nonderogatory. This is why the generic matrix has an $n$-dimensional centraliser while the generic nilpotent, at $n=3$, has a five-dimensional one.

The Companion Matrix

Every monic polynomial is the characteristic polynomial of a companion matrix, and over $\mathbb{R}$ the companion matrix is real. For $p(t)=t^{n}+c_{n-1}t^{n-1}+\cdots+c_0$ put

$$ C_p=\begin{pmatrix}0&0&\cdots&0&-c_0\\ 1&0&\cdots&0&-c_1\\ 0&1&\cdots&0&-c_2\\ \vdots&&\ddots&&\vdots\\ 0&0&\cdots&1&-c_{n-1}\end{pmatrix}. $$

Then $\chi_{C_p}=p=m_{C_p}$, so the companion matrix is nonderogatory and realises every real polynomial as a minimal polynomial.

The companion matrix is the multiplication operator on the quotient ring. On $\mathbb{R}[t]/(p)$ of dimension $n$, multiplication by $t$ has matrix $C_p$ in the basis $1,t,\dots,t^{n-1}$. This is the cleanest proof that $\chi_{C_p}=p$ and that the algebra generated is the whole quotient.

Consequence: every conjugacy class over $\mathbb{R}$ is represented by a rational canonical form. Assembling companion matrices of the invariant factors, that is of the elementary divisors of $A$ over $\mathbb{R}[t]$, gives a canonical form with entries in $\mathbb{R}$ and no rotation block written separately. The real Jordan form is the same data in the eigenbasis; the rational canonical form is the same data in a basis of cyclic vectors.

Summary

The minimal polynomial is the monic generator of the annihilator, it divides the characteristic polynomial, and its roots are the spectrum with exponents the sizes of the largest blocks. Over $\mathbb{R}$ its irreducible factors are linear or quadratic; a quadratic factor is a rotation block. The real Jordan form is a direct sum of blocks $J_d(\lambda)$ for real eigenvalues and $R_d(a,b)$ for conjugate pairs, and the list is unique. The centraliser has dimension the sum of the squares of the block multiplicities, and it equals the algebra generated exactly for the nonderogatory matrices. Every real polynomial is the minimal polynomial of a real companion matrix.

Summary of Notation

Symbol Meaning
$\operatorname{Ann}(A)$ the ideal of polynomials annihilating $A$
$m_A$ the minimal polynomial, the monic generator
$J_d(\lambda)$ the Jordan block of size $d$ for the real eigenvalue $\lambda$
$R_d(a,b)$ the real rotation block of size $2d$ for the pair $a\pm bi$
$Z(A)$ the centraliser of $A$
$C_p$ the companion matrix of the polynomial $p$

Further Reading