The Minimal Polynomial and the Jordan Form
Introduction
The characteristic polynomial gives the spectrum with multiplicities, the minimal polynomial gives the spectrum with the size of the largest block, and the Jordan form gives everything: the two polynomials are two summaries of one combinatorial datum, the partition of $n$ attached to each eigenvalue.
The article defines the minimal polynomial as the monic generator of the annihilator of $A$, proves that it divides the characteristic polynomial and has the same roots, constructs the Jordan form from the generalised eigenspaces, and states the classification: two matrices are conjugate exactly when their Jordan types agree. It then counts: the number of types for a nilpotent of size $n$ is the number of partitions of $n$, and the dimension of the centraliser is the sum of the squares of the conjugate partition. It assumes The Element: Rank, Trace, Determinant, Characteristic Polynomial for the invariants and The Elementary Symmetric Functions and $\det(\mathrm{I}+tA)$ for the coefficients. It defers the idempotents that the blocks produce to The Peirce Decomposition and the Two Halves and the real classification to The Two Base Fields: $M_n(\mathbb{R})$ inside $M_n(\mathbb{C})$.
The Annihilator
The polynomials that kill $A$ form a nonzero ideal, and its monic generator is the minimal polynomial. Put $\mathcal{I}(A)=\{f\in\mathbb{C}[t]: f(A)=0\}$. It is an ideal of $\mathbb{C}[t]$, and it is nonzero because the $n^{2}+1$ powers $\mathrm{I},A,\dots,A^{n^{2}}$ lie in the $n^{2}$-dimensional space $M_n(\mathbb{C})$ and are therefore dependent. The ring $\mathbb{C}[t]$ is principal, so $\mathcal{I}(A)=(m_A)$ with $m_A$ monic, and $m_A$ is the minimal polynomial of $A$.
The minimal polynomial is a conjugacy invariant of degree at most $n$. For $S$ invertible, $f(SAS^{-1})=Sf(A)S^{-1}$, so $\mathcal{I}(SAS^{-1})=\mathcal{I}(A)$. The degree bound is Cayley–Hamilton: $\chi_A(A)=0$, so $\chi_A\in\mathcal{I}(A)$ and
$$ m_A\mid\chi_A,\qquad \deg m_A\le n . $$
The degree of $m_A$ is the dimension of the algebra $\mathbb{C}[A]$, since the powers of $A$ are independent exactly up to that degree.
The minimal polynomial and the characteristic polynomial have the same roots. If $Av=\lambda v$ with $v\ne0$, then $0=m_A(A)v=m_A(\lambda)v$, so $m_A(\lambda)=0$. Conversely, if $m_A(\lambda)=0$, write $m_A=(t-\lambda)g$; by minimality $g(A)\ne0$, so there is $v$ with $w=g(A)v\ne0$, and $(A-\lambda)w=m_A(A)v=0$. So $\lambda$ is an eigenvalue. Therefore
$$ m_A(t)=\prod_{i=1}^{r}(t-\lambda_i)^{a_i},\qquad \{m_A=0\}=\{\chi_A=0\}, $$
with the $\lambda_i$ distinct, and the exponent $a_i$ is the size of the largest Jordan block attached to $\lambda_i$.
The matrix is nonderogatory exactly when the two polynomials agree. The equality $m_A=\chi_A$ means one block per eigenvalue, and then $\mathbb{C}[A]$ has dimension $n$ and is a maximal commutative subalgebra. The generic matrix is nonderogatory, since the condition that two eigenvalues coalesce or that a block splits is the vanishing of polynomial conditions.
The Jordan Form
Every matrix is conjugate to a direct sum of Jordan blocks, uniquely up to the order of the summands. A Jordan block is
$$ J_d(\lambda)=\lambda\mathrm{I}+N_d,\qquad N_d=\sum_{k=1}^{d-1}E_{k,k+1}, $$
with $N_d$ nilpotent of index $d$. The theorem states that for every $A$ there is an invertible $S$ and a list of pairs $(d_i,\lambda_i)$ with
$$ A=S\Big(\bigoplus_{i}J_{d_i}(\lambda_i)\Big)S^{-1}, $$
and that the list is unique once the blocks are grouped by eigenvalue and sorted by size.
The construction is the one of the generalised eigenspaces. Let $V_\lambda=\ker(A-\lambda\mathrm{I})^{n}$, the generalised eigenspace. The space $V$ is the direct sum of the $V_\lambda$, and on each summand $A-\lambda\mathrm{I}$ is nilpotent. A nilpotent is a sum of a chain of nested subspaces, and choosing a basis adapted to the chain exhibits the block. The sizes of the blocks are the numerical invariants of the nilpotent, read from the ranks
$$ \dim\ker N^{k}-\dim\ker N^{k-1}=\#\{i: d_i\ge k\}, $$
so the ranks of the powers determine the partition and nothing else does.
The rank of a nilpotent counts the blocks. For a nilpotent with block sizes $d_1,\dots,d_r$ and $d_1+\cdots+d_r=n$,
$$ \operatorname{rank}N=n-r . $$
So the identity $N=0$ has rank $0$ and $n$ blocks, and a single block of size $n$ has rank $n-1$. This is the sharpest illustration of the failure recorded in The Element: Rank, Trace, Determinant, Characteristic Polynomial: the characteristic polynomial of every nilpotent is $t^{n}$, and the rank separates the types.
The Classification
Two matrices are conjugate exactly when their Jordan types agree. Conjugation preserves the list of blocks, because it preserves the generalised eigenspaces and the ranks of the powers of $A-\lambda\mathrm{I}$. Conversely, two matrices with the same list are conjugate, since each is conjugate to the same direct sum. Hence
$$ A\sim B\iff (d_i,\lambda_i)_A=(d_i,\lambda_i)_B\ \text{up to order}. $$
The classification is the orbit decomposition of the conjugation action. Each Jordan type is one orbit of $GL_n(\mathbb{C})$ on $M_n(\mathbb{C})$, and the orbit of a nonderogatory matrix is open and dense: its dimension is $n^{2}-n+1$, the dimension of $GL_n$ minus the dimension $n-1$ of the centraliser of the companion matrix. The union of the other orbits is the discriminant, the zero set of the polynomial that says two eigenvalues coincide or a block splits.
The Jordan form is a refinement of the characteristic polynomial and a coarsening of the element. For a fixed characteristic polynomial $\prod(t-\lambda_i)^{m_i}$ the types are the tuples of partitions of the multiplicities $m_i$. The number of types with a given spectrum is therefore the product of the partition numbers of the multiplicities, and the spectrum is recovered from the type.
The Centraliser and the Commutant
The centraliser of a matrix is the space of matrices commuting with it, and its dimension is read from the blocks. For $A$ with blocks of sizes $d_{i,1}\ge d_{i,2}\ge\cdots$ attached to the eigenvalue $\lambda_i$, let $\lambda^{(i)}$ be the conjugate partition of the sizes. Then
$$ \dim C(A)=\sum_{i}\sum_{j}\big(\lambda^{(i)}_{j}\big)^{2}. $$
The count is verified on the nilpotents of size at most four: the blocks $J_4,J_3\oplus J_1,J_2\oplus J_2,J_2\oplus J_1\oplus J_1,0$ give $4,6,8,10,16$, and the zero matrix gives $16=n^{2}$ as it must.
The algebra generated by $A$ sits inside the centraliser, and the inclusion is strict except in the nonderogatory case. Since $\mathbb{C}[A]$ is commutative and contains $A$, $\mathbb{C}[A]\subseteq C(A)$, and the dimensions are $\deg m_A$ and the number above. For the zero matrix the left side is $1$ and the right side is $n^{2}$. The equality $\mathbb{C}[A]=C(A)$ holds exactly when $m_A=\chi_A$.
The centraliser of a single element is a product of matrix algebras, one for each eigenvalue. On the block of $\lambda_i$ the centraliser is the algebra of matrices that are upper triangular in the flag of the chain and constant on each diagonal of the chain; it is a quotient of a product of matrix algebras, and its dimension is the sum of the squares above. The case of a single block of size $n$ is the extreme: $C(J_n(\lambda))=\mathbb{C}[J_n]$, of dimension $n$, the polynomials in one nilpotent.
The Number of Types at Small $n$
The nilpotent types of size $n$ are the partitions of $n$. The counts are
| $n$ | $2$ | $3$ | $4$ | $5$ | $6$ |
|---|---|---|---|---|---|
| partitions of $n$ | $2$ | $3$ | $5$ | $7$ | $11$ |
and the lists are
$$ (2),\ (1,1);\qquad (3),\ (2,1),\ (1,1,1);\qquad (4),\ (3,1),\ (2,2),\ (2,1,1),\ (1,1,1,1). $$
Each partition is one conjugacy class of nilpotents, and the classes are ordered by the rank $n-r$ and by the index of nilpotency, the largest part.
The total number of types at size $n$ is infinite for $n\ge1$, since the eigenvalues are free, but the combinatorial part is finite. For a fixed spectrum the number of types is the product of the partition numbers of the multiplicities; for a single eigenvalue of multiplicity $n$ it is the partition number of $n$. So at $n=6$ a matrix with one eigenvalue has $11$ possible types, and the generic matrix, with six distinct eigenvalues, has exactly one.
The Peirce decomposition sees the blocks, not the types. A complete family of $n$ idempotents requires $n$ distinct spectral lines, hence $n$ distinct eigenvalues and blocks of size one. A single block of size $n$ carries one eigenvalue and one idempotent, and the construction of The Peirce Decomposition in $n$ Dimensions degenerates to the trivial one. The passage from the Jordan form to the Peirce frame is the passage from one eigenvalue to a full spectrum.
The Case $n=2$
At $n=2$ there are three kinds and the discriminant separates them. For $A\in M_2(\mathbb{C})$ with $\chi_A(t)=t^{2}-\tau t+\delta$, where $\tau=\operatorname{tr}A$ and $\delta=\det A$:
| condition | type | minimal polynomial |
|---|---|---|
| $\Delta=\tau^{2}-4\delta\ne0$ | two distinct eigenvalues, diagonalisable | $(t-\lambda_1)(t-\lambda_2)$ |
| $\Delta=0$, $A\ne\tfrac{\tau}{2}\mathrm{I}$ | one eigenvalue, one block of size two | $(t-\tfrac{\tau}{2})^{2}$ |
| $A=\tfrac{\tau}{2}\mathrm{I}$ | scalar | $t-\tfrac{\tau}{2}$ |
The middle row is the interesting one: the minimal polynomial has degree two and equals the characteristic polynomial, and $A-\tfrac{\tau}{2}\mathrm{I}$ is a nonzero nilpotent with rank $1$.
The middle row is the companion nilpotent of the biquaternion frame. At $n=2$ a nonzero nilpotent $\vec\Upsilon_{+-}$ of The Companion Nilpotents of the Biquaternion Algebra satisfies $\vec\Upsilon_{+-}^{2}=0$ and has rank $1$, so its minimal polynomial is $t^{2}$ and its Jordan type is $(2)$. The two companion nilpotents $\vec\Upsilon_{+-}$ and $\vec\Upsilon_{-+}$ are conjugate, since there is one type, and they are distinguished only by the frame, which is the gauge freedom of The Frame, the Simple Nilpotents and the Gauge of $M_n(\mathbb{C})$.
The classification at $n=2$ is the classification of the real forms at $n=2$ read differently. A real $2\times2$ matrix with $\Delta<0$ is not diagonalisable over $\mathbb{R}$ and is a rotation block, which is the real normal form of The Two Base Fields: $M_n(\mathbb{R})$ inside $M_n(\mathbb{C})$; the complex type is two distinct eigenvalues and the real type is one $2\times2$ block. The two classifications differ in the field and not in the algebra.
Summary
The minimal polynomial $m_A$ is the monic generator of the ideal of polynomials killing $A$; it divides $\chi_A$, it has degree at most $n$, and it has the same roots as $\chi_A$, with exponents equal to the sizes of the largest blocks. The Jordan form writes $A$ as a direct sum of blocks $J_d(\lambda)$, uniquely up to order, and two matrices are conjugate exactly when their lists agree. For a nilpotent the types are the partitions of $n$, counted $2,3,5,7,11$ for $n=2,\dots,6$, the rank is $n$ minus the number of blocks, and the centraliser has dimension the sum of the squares of the conjugate partition. At $n=2$ the three kinds are separated by the discriminant, and the non-scalar double-eigenvalue case is the companion nilpotent of the biquaternion frame.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $\mathcal{I}(A)$ | the ideal of polynomials $f$ with $f(A)=0$ |
| $m_A$ | the minimal polynomial, the monic generator of $\mathcal{I}(A)$ |
| $\chi_A$ | the characteristic polynomial |
| $J_d(\lambda)$ | the Jordan block of size $d$ with eigenvalue $\lambda$ |
| $N_d$ | the nilpotent $\sum_{k |
| $V_\lambda=\ker(A-\lambda\mathrm{I})^{n}$ | the generalised eigenspace |
| $C(A)$ | the centraliser, $\{X: AX=XA\}$ |
| $\lambda^{(i)}$ | the conjugate partition of the block sizes at $\lambda_i$ |
| $\Delta=(\operatorname{tr}A)^{2}-4\det A$ | the discriminant at $n=2$ |
Further Reading
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge, 2013), for the Jordan form, the minimal polynomial and the centraliser.
- James E. Humphreys, Linear Algebraic Groups (Springer, 1975), for the orbit decomposition of the conjugation action and the nonderogatory open orbit.
- Peter Lancaster and Miron Tismenetsky, The Theory of Matrices (Academic Press, 1985), for the structure of the commutant of a single matrix.
- The Element: Rank, Trace, Determinant, Characteristic Polynomial, for the invariants that the type refines.
- The Elementary Symmetric Functions and $\det(\mathrm{I}+tA)$, for the coefficients of $\chi_A$ and Newton's identities.
- The Peirce Decomposition in $n$ Dimensions, for the full spectrum that the frame requires.
- The Companion Nilpotents of the Biquaternion Algebra, for the two nilpotents at $n=2$.