9.3 The Jordan Canonical Form
Every complex square matrix is similar to an essentially unique block-diagonal matrix of Jordan blocks: the ranks of $(A-\lambda I)^k$ read it off, partitions count the similarity classes, and the rational canonical form does the same job over any field.
Diagonalisation (§9.2) is the ideal outcome: in a basis of eigenvectors an operator is just a list of scalings. But many operators have too few eigenvectors — the shear $\begin{pmatrix}1&1\\0&1\end{pmatrix}$, differentiation on polynomials, every nonzero nilpotent matrix — and no change of basis will make them diagonal. The Jordan canonical form is the best that can be done instead. Over $\mathbb{C}$, every operator is a direct sum of pieces of one standard shape, $\lambda I+N$ with $N$ the "shift" that sends each basis vector to the previous one, and the sizes of those pieces are a complete similarity invariant: two complex matrices are similar exactly when they have the same Jordan blocks. The whole classification of matrices up to similarity therefore comes down to a little combinatorics — partitions of integers — read off from the ranks of the powers $(A-\lambda I)^k$.
The Jordan form is how one actually computes $A^m$ and $e^{tA}$ for defective matrices (the source of the $te^{\lambda t}$ terms in the solutions of linear ODEs in Unit 3), decides when $A^m\to0$, and proves facts such as "every matrix is similar to its transpose". It is also among the most examined topics of CSIR NET linear algebra: "the number of similarity classes of $6\times6$ matrices with characteristic polynomial … and minimal polynomial …", "if $\operatorname{rank}(A-2I)=3$ and $\operatorname{rank}(A-2I)^2=1$, how many Jordan blocks …", "which of these nilpotent matrices are similar" appear in some form in almost every paper. This section develops the theory with proofs — generalised eigenvectors, the structure of nilpotent operators, existence and uniqueness of the Jordan form — and then the counting techniques and the rational canonical form that answer those questions in a few lines.
Throughout, $V$ is a finite-dimensional vector space over a field $\mathbb{F}$, $n=\dim V\ge1$, and $T\in\mathcal{L}(V)$; for a matrix $A\in M_n(\mathbb{F})$ we apply everything to $x\mapsto Ax$. As in §9.1, $p_T(x)=\det(xI-T)$ is the characteristic polynomial and $m_T$ the minimal polynomial; for an eigenvalue $\lambda$, $a_\lambda$ is its algebraic and $g_\lambda=\dim\ker(T-\lambda I)$ its geometric multiplicity. We use the facts from §9.1 that $m_T$ divides every polynomial that annihilates $T$, that the roots of $m_T$ are exactly the eigenvalues, and the Cayley–Hamilton theorem $p_T(T)=0$.
Generalised Eigenvectors and the Fitting Decomposition
If $A=\begin{pmatrix}2&1\\0&2\end{pmatrix}$, then $e_2$ is not an eigenvector, but $(A-2I)e_2=e_1$ is, so $(A-2I)^2e_2=0$. Vectors killed by a power of $A-\lambda I$ are the missing building blocks.
Let $\lambda\in\mathbb{F}$ and $N=T-\lambda I$. A nonzero vector $v\in V$ is a generalised eigenvector of $T$ for $\lambda$ if $N^kv=0$ for some $k\ge1$; the least such $k$ is the height of $v$ (some books say "rank" or "order"). Eigenvectors are exactly the generalised eigenvectors of height $1$. The generalised eigenspace of $\lambda$ is
$$G_\lambda(T)=\{v\in V:\ (T-\lambda I)^kv=0\ \text{for some }k\ge1\}=\bigcup_{k\ge1}\ker(T-\lambda I)^k.$$- Stabilisation. $\ker N\subseteq\ker N^2\subseteq\cdots$, and there is an integer $s$ with $0\le s\le n$ such that the inclusions are strict up to $\ker N^s$ and $\ker N^s=\ker N^{s+1}=\ker N^{s+2}=\cdots$. Hence $G_\lambda(T)=\ker N^s=\ker N^n$ is a subspace.
- Fitting decomposition. $V=\ker N^s\oplus\operatorname{im}N^s$, and both summands are $T$-invariant.
- $N$ restricted to $G_\lambda(T)$ is nilpotent of index $s$, and $N$ restricted to $\operatorname{im}N^s$ is invertible. For every $\mu\ne\lambda$, $T-\mu I$ maps $G_\lambda(T)$ bijectively onto itself. Consequently $G_\lambda(T)\ne\{0\}$ iff $\lambda$ is an eigenvalue, and $G_\lambda(T)\cap G_\mu(T)=\{0\}$ for $\lambda\ne\mu$.
- If $\lambda$ is an eigenvalue, $s$ equals the multiplicity of $\lambda$ as a root of $m_T$.
Proof. (1) $N^jv=0$ implies $N^{j+1}v=0$, so the kernels increase. Suppose $\ker N^j=\ker N^{j+1}$ for some $j$. If $N^{j+2}v=0$ then $Nv\in\ker N^{j+1}=\ker N^j$, so $N^{j+1}v=0$: thus $\ker N^{j+1}=\ker N^{j+2}$, and by induction all later kernels coincide. So the chain increases strictly up to some first $s$ and is constant afterwards; each strict step raises the dimension by at least $1$ and the dimensions are at most $n$, so $s\le n$. Then $\bigcup_k\ker N^k=\ker N^s=\ker N^n$.
(2) $T$ commutes with $N=T-\lambda I$, hence with $N^s$; so if $N^sv=0$ then $N^s(Tv)=T(N^sv)=0$, and $T(N^sw)=N^s(Tw)$: both $\ker N^s$ and $\operatorname{im}N^s$ are $T$-invariant. If $v\in\ker N^s\cap\operatorname{im}N^s$, write $v=N^sw$; then $N^{2s}w=0$, so $w\in\ker N^{2s}=\ker N^s$ and $v=N^sw=0$. By rank–nullity $\dim\ker N^s+\dim\operatorname{im}N^s=n$, so the sum is all of $V$.
(3) $N^s=0$ on $G_\lambda=\ker N^s$, while $N^{s-1}\ne0$ there because $\ker N^{s-1}\ne\ker N^s$; so the index is $s$. On $W=\operatorname{im}N^s$, the kernel of $N$ is $\ker N\cap W\subseteq\ker N^s\cap W=\{0\}$, so $N|_W$ is injective, hence invertible ($W$ is finite-dimensional and $N$-invariant). For $\mu\ne\lambda$, on $G_\lambda$ we have $T-\mu I=(\lambda-\mu)\big(I+(\lambda-\mu)^{-1}N\big)$, and $I+cN$ is invertible when $N$ is nilpotent — its inverse is $I-cN+c^2N^2-\cdots$, a finite sum (the structure theorem for nilpotent maps, §8.2). Finally, $G_\lambda\ne\{0\}$ iff $\ker N\ne\{0\}$ (if $\ker N=\{0\}$ then every $N^k$ is injective), and if $v\in G_\lambda\cap G_\mu$ then $(T-\mu I)^kv=0$ for some $k$, while $(T-\mu I)^k$ is injective on the $T$-invariant space $G_\lambda$; so $v=0$.
(4) Write $m_T(x)=(x-\lambda)^eq(x)$ with $q(\lambda)\ne0$. On $G_\lambda$, $q(T)=q(\lambda I+N)=q(\lambda)I+N\,r(N)$ for some polynomial $r$, and $N\,r(N)$ is nilpotent, so $q(T)$ is invertible on $G_\lambda$. For $v\in G_\lambda$: $0=m_T(T)v=q(T)\big(N^ev\big)$ with $N^ev\in G_\lambda$, hence $N^ev=0$. So $N^e=0$ on $G_\lambda$, and $s\le e$. Conversely, let $W=\operatorname{im}N^s$. Since $N$ is invertible on $W$, every $w\in W$ is $N^ew'$ with $w'\in W$, and $q(T)w=q(T)N^ew'=m_T(T)w'=0$: $q(T)$ vanishes on $W$. Put $g(x)=(x-\lambda)^sq(x)$. Then $g(T)=N^sq(T)=q(T)N^s$ kills $\ker N^s$ (via $N^s$) and kills $W$ (via $q(T)$), so $g(T)=0$ on $V=\ker N^s\oplus W$. Therefore $m_T\mid g$, which gives $e\le s$. ∎
- $V=G_1\oplus G_2\oplus\cdots\oplus G_r$, each $G_i$ being $T$-invariant;
- $\dim G_i=a_i$, and the restriction $T_i=T|_{G_i}$ has $p_{T_i}(x)=(x-\lambda_i)^{a_i}$;
- $m_T(x)=\prod_{i=1}^r(x-\lambda_i)^{s_i}$, with $1\le s_i\le a_i$.
Proof. Induction on $r$. Let $\lambda=\lambda_1$, $N=T-\lambda I$ and $W=\operatorname{im}N^{s_1}$. By Theorem 9.21(2), $V=G_1\oplus W$ with both parts $T$-invariant, so in an adapted basis $[T]=\operatorname{diag}(A_1,A_2)$ and $p_T=p_{T|_{G_1}}\cdot p_{T|_W}$. On $G_1$, $T=\lambda I+N$ with $N$ nilpotent; choose a basis of $G_1$ by extending a basis of $\ker N$ to one of $\ker N^2$, then of $\ker N^3$, and so on. Since $N$ maps $\ker N^{j}$ into $\ker N^{j-1}$, the matrix of $N|_{G_1}$ in this basis is block upper triangular with zero diagonal blocks, so $p_{T|_{G_1}}(x)=(x-\lambda)^{d}$, $d=\dim G_1$. On $W$, $T-\lambda I$ is invertible (Theorem 9.21(3)), so $\lambda$ is not a root of $p_{T|_W}$. Comparing the two factorisations of $p_T$ gives $d=a_1$ and $p_{T|_W}=\prod_{i\ge2}(x-\lambda_i)^{a_i}$. If $r=1$ this says $\dim W=0$ and we are done. Otherwise $T|_W$ has split characteristic polynomial with $r-1$ distinct roots, and by induction $W=\bigoplus_{i\ge2}G_{\lambda_i}(T|_W)$ with $\dim G_{\lambda_i}(T|_W)=a_i$.
It remains to see that $G_{\lambda_i}(T|_W)=G_{\lambda_i}(T)$ for $i\ge2$. The inclusion $\subseteq$ is clear. If $v\in G_{\lambda_i}(T)$, write $v=g+w$ with $g\in G_1$, $w\in W$; then $0=(T-\lambda_iI)^nv=(T-\lambda_iI)^ng+(T-\lambda_iI)^nw$, the two terms lying in the invariant subspaces $G_1$ and $W$. Hence $(T-\lambda_iI)^ng=0$, i.e. $g\in G_1\cap G_{\lambda_i}=\{0\}$, and $v=w\in G_{\lambda_i}(T|_W)$. This proves (1) and (2). For (3): $m_T$ divides $p_T$ (Cayley–Hamilton), so $m_T=\prod(x-\lambda_i)^{e_i}$, every $e_i\ge1$ because each $\lambda_i$ is a root of $m_T$ (§9.1), and $e_i=s_i$ by Theorem 9.21(4); finally $s_i\le\dim G_i=a_i$ since the kernels $\ker N^j$ inside $G_i$ grow strictly for $j\le s_i$. ∎
Theorem 9.22 is the primary decomposition theorem of §9.2 in the case of a split characteristic polynomial (there the projections onto the $G_i$ are shown to be polynomials in $T$). The new information is part (3) together with Theorem 9.21(4): the exponent of $(x-\lambda)$ in $m_T$ is the step at which the kernels of $(T-\lambda I)^k$ stop growing. What is left is to understand a single nilpotent operator $N_i$ — that is the content of the next subsection.
Jordan Blocks and Nilpotent Operators
For $\lambda\in\mathbb{F}$ and $k\ge1$ the Jordan block $J_k(\lambda)$ is the $k\times k$ matrix with $\lambda$ on the diagonal, $1$ on the superdiagonal and $0$ elsewhere: $$J_k(\lambda)=\begin{pmatrix}\lambda&1&&\\&\lambda&\ddots&\\&&\ddots&1\\&&&\lambda\end{pmatrix},\qquad J_1(\lambda)=(\lambda).$$ A Jordan matrix is a block-diagonal matrix $\operatorname{diag}\big(J_{k_1}(\lambda_1),\dots,J_{k_m}(\lambda_m)\big)$; the $\lambda_i$ need not be distinct.
A Jordan chain of length $k$ for $T$ and $\lambda$ is a list $v_1,\dots,v_k$ of nonzero vectors with $(T-\lambda I)v_1=0$ and $(T-\lambda I)v_j=v_{j-1}$ for $2\le j\le k$; equivalently $v_j=(T-\lambda I)^{k-j}v_k$ where $v_k$ has height $k$. The bottom $v_1$ is an eigenvector, the top $v_k$ generates the chain. A Jordan basis is a basis of $V$ that is a disjoint union of Jordan chains; a Jordan form of $T$ is $[T]_{\mathcal{B}}$ for a Jordan basis $\mathcal{B}$ (each chain, listed bottom first, contributes one block).
- $p_J(x)=m_J(x)=(x-\lambda)^k$.
- The eigenspace of $\lambda$ is $\operatorname{span}\{e_1\}$; so $g_\lambda=1$ while $a_\lambda=k$.
- $\operatorname{rank}(J-\lambda I)^j=\max(k-j,\,0)$ for every $j\ge0$.
- For $m\ge0$, $\displaystyle J^m=\sum_{j=0}^{k-1}\binom{m}{j}\lambda^{m-j}N^j$: the entry of $J^m$ in position $(i,i+j)$ is $\binom{m}{j}\lambda^{m-j}$ (with $\binom{m}{j}=0$ for $j\gt m$).
- $J$ is invertible iff $\lambda\ne0$; $J\sim J^{T}$; and $J_k(0)\sim c\,J_k(0)$ for every $c\ne0$.
Proof. (3) $N^je_i=e_{i-j}$ if $i\gt j$ and $N^je_i=0$ otherwise, so $N^j$ has ones exactly on the $j$-th superdiagonal: rank $k-j$ for $j\lt k$, and $N^k=0$. (2) is the case $j=1$: $\ker N=\operatorname{span}\{e_1\}$. (1) $J$ is upper triangular with diagonal $\lambda$, so $p_J=(x-\lambda)^k$; $m_J$ divides it and $(J-\lambda I)^{k-1}=N^{k-1}\ne0$, so $m_J=(x-\lambda)^k$. (4) $\lambda I$ and $N$ commute, so the binomial theorem applies, and $N^j=0$ for $j\ge k$. (5) $\det J=\lambda^k$. Let $R$ be the reversal permutation matrix, $Re_i=e_{k+1-i}$, so $R=R^{-1}$; then $(RJR)_{ij}=J_{k+1-i,\,k+1-j}$, which is $\lambda$ for $i=j$ and $1$ exactly when $j=i-1$: $RJR=J^{T}$. For the last claim let $S=\operatorname{diag}(1,c,c^2,\dots,c^{k-1})$; the $(i,i+1)$ entry of $S^{-1}NS$ is $c^{i}/c^{i-1}=c$, so $S^{-1}NS=cN$. ∎
By (3), a Jordan block is recognised by ranks: $J_k(\lambda)-\lambda I$ loses exactly one rank with each further power until it reaches $0$ at the $k$-th. The structure theorem below shows that every nilpotent operator is a direct sum of such shifts, and that the rank sequence tells us how many shifts of each length there are. Its statement uses a little combinatorics.
- Existence. $V$ has a Jordan basis for $N$ (eigenvalue $0$): $N\sim\operatorname{diag}\big(J_{k_1}(0),\dots,J_{k_t}(0)\big)$ with $k_1\ge\cdots\ge k_t\ge1$.
- Uniqueness. For every $j\ge1$, the number of blocks of size $\ge j$ is $$d_j=\dim\ker N^j-\dim\ker N^{j-1}=\operatorname{rank}N^{j-1}-\operatorname{rank}N^j,$$ and the number of blocks of size exactly $j$ is $d_j-d_{j+1}$ $=\operatorname{rank}N^{j-1}-2\operatorname{rank}N^j+\operatorname{rank}N^{j+1}$. So the Jordan type is determined by $N$, and its conjugate is the sequence of kernel jumps.
- The number of blocks is $t=\dim\ker N$, and the largest block $k_1$ is the index of nilpotency.
- Two nilpotent $n\times n$ matrices are similar iff they have the same Jordan type. Hence the similarity classes of nilpotent $n\times n$ matrices correspond bijectively to the partitions of $n$, and there are $p(n)$ of them.
Proof. (1) Strong induction on $n$. If $N=0$ (in particular if $n=1$), any basis consists of chains of length $1$. Suppose $N\ne0$ and let $U=\operatorname{im}N$. It is $N$-invariant, $N|_U$ is nilpotent, $U\ne\{0\}$, and $\dim U\lt n$ because a nilpotent map has a nonzero kernel. By the induction hypothesis $U$ has a Jordan basis for $N|_U$, consisting of $r$ chains which we write top-down as $$u_i,\ Nu_i,\ \dots,\ N^{\ell_i-1}u_i\qquad(N^{\ell_i}u_i=0,\ 1\le i\le r).$$ Each $u_i$ lies in $\operatorname{im}N$, so we may choose $w_i\in V$ with $Nw_i=u_i$. The bottoms $N^{\ell_i-1}u_i=N^{\ell_i}w_i$ lie in $\ker N\cap U$ and are linearly independent; in fact they form a basis of $\ker N\cap U$, because $N$ maps the chain basis of $U$ onto the other basis vectors (the bottoms go to $0$), so $\operatorname{rank}N|_U=\dim U-r$ and $\dim(\ker N\cap U)=r$. Extend the bottoms to a basis of $\ker N$ by vectors $z_1,\dots,z_q$. We claim that $$\mathcal{B}=\{N^jw_i:\ 0\le j\le\ell_i,\ 1\le i\le r\}\cup\{z_1,\dots,z_q\}$$ is a basis of $V$; it is visibly a union of chains (of lengths $\ell_i+1$, and $q$ chains of length $1$). It has $\sum_i(\ell_i+1)+q$ $=\dim U+r+q$ $=\operatorname{rank}N+\dim\ker N=n$ elements, so it suffices to prove independence. Suppose $\sum_{i,j}c_{ij}N^jw_i+\sum_lb_lz_l=0$. Applying $N$ kills the $z_l$ and the top-index terms $N^{\ell_i}w_i$, and turns $N^jw_i$ into $N^{j}u_i$ for $j\lt\ell_i$: we get $\sum_i\sum_{j\lt\ell_i}c_{ij}N^ju_i=0$, a relation among the basis vectors of $U$, so $c_{ij}=0$ for $j\lt\ell_i$. What is left, $\sum_ic_{i\ell_i}N^{\ell_i}w_i+\sum_lb_lz_l=0$, is a relation in the chosen basis of $\ker N$, so all remaining coefficients vanish.
(2) Rank is a similarity invariant (§8.2), so we may compute in the Jordan matrix $\operatorname{diag}(J_{k_1}(0),\dots,J_{k_t}(0))$. By Theorem 9.23(3) the block $J_{k_i}(0)$ contributes $\max(k_i-j,0)$ to $\operatorname{rank}N^j$, so $$\operatorname{rank}N^{j-1}-\operatorname{rank}N^j=\sum_{i=1}^t\big[\max(k_i-j+1,0)-\max(k_i-j,0)\big]=\#\{i:\ k_i\ge j\},$$ each bracket being $1$ if $k_i\ge j$ and $0$ otherwise. The numbers $d_j$ depend only on $N$, and they determine how many $k_i$ equal each value; so every Jordan basis of $N$ has the same chain lengths. The identity $\dim\ker N^j-\dim\ker N^{j-1}$ $=\operatorname{rank}N^{j-1}-\operatorname{rank}N^j$ is rank–nullity. (3) $d_1=n-\operatorname{rank}N=\dim\ker N$; and $N^j=0$ iff every $k_i\le j$. (4) Similar matrices have the same ranks of powers, hence the same type by (2); matrices of the same type are similar to the same Jordan matrix. Every partition occurs (take the Jordan matrix itself). ∎
In the dot diagram the proof is a picture: row $i$ is a chain, the leftmost dot its eigenvector, and $N$ moves every dot one place to the left (the leftmost to $0$). So $\ker N^j$ is spanned by the first $j$ columns of dots, and the $j$-th column has $d_j=\dim\ker N^j-\dim\ker N^{j-1}$ dots. The inequality $d_1\ge d_2\ge\cdots$ — the kernel jumps never increase — is the structure theorem for nilpotent maps of §8.2 seen again.
The Jordan Canonical Form: Existence, Uniqueness and Computation
- Existence. $V$ has a Jordan basis for $T$: there is a basis $\mathcal{B}$ with $[T]_{\mathcal{B}}=\operatorname{diag}\big(J_{k_1}(\lambda_1),\dots,J_{k_m}(\lambda_m)\big)$.
- Uniqueness. The Jordan form is unique up to the order of the blocks: for each $\lambda$ and $k\ge1$, writing $r_j(\lambda)=\operatorname{rank}(T-\lambda I)^j$ (so $r_0=n$), the number of blocks equal to $J_k(\lambda)$ is $$r_{k-1}(\lambda)-2r_k(\lambda)+r_{k+1}(\lambda),$$ and the number of blocks $J_j(\lambda)$ with $j\ge k$ is $r_{k-1}(\lambda)-r_k(\lambda)$.
- Two matrices $A,B\in M_n(\mathbb{C})$ are similar iff they have the same Jordan form, i.e. iff $\operatorname{rank}(A-\lambda I)^k=\operatorname{rank}(B-\lambda I)^k$ for every $\lambda\in\mathbb{C}$ and every $k\ge1$.
Proof. (1) By Theorem 9.22, $V=G_1\oplus\cdots\oplus G_r$ with $T_i=\lambda_iI+N_i$ on $G_i$ and $N_i$ nilpotent. By Theorem 9.24, $G_i$ has a Jordan basis for $N_i$; a Jordan chain for $N_i$ and $0$ is a Jordan chain for $T$ and $\lambda_i$, because $T-\lambda_iI=N_i$ on $G_i$. The union of these bases is a Jordan basis of $V$.
(2) Let $J=[T]_{\mathcal{B}}$ be any Jordan form of $T$. Then $(J-\lambda I)^j$ is block diagonal, so its rank is the sum of the ranks of $(J_k(\mu)-\lambda I)^j$ over the blocks. If $\mu\ne\lambda$, $J_k(\mu)-\lambda I$ is triangular with nonzero diagonal $\mu-\lambda$, hence invertible, and contributes $k$ for every $j$. If $\mu=\lambda$ it contributes $\max(k-j,0)$ (Theorem 9.23(3)). Subtracting, the blocks with $\mu\ne\lambda$ cancel and, exactly as in the proof of Theorem 9.24(2), $r_{k-1}(\lambda)-r_k(\lambda)$ counts the blocks $J_j(\lambda)$ with $j\ge k$. The second difference counts those with $j=k$. Since $r_j(\lambda)=\operatorname{rank}(T-\lambda I)^j$ does not depend on the basis, all Jordan forms of $T$ contain the same blocks with the same multiplicities.
(3) Similar matrices are the matrices of one operator in two bases (§8.2), so they have the same Jordan form by (2), and the same ranks of $(A-\lambda I)^k$. Conversely, if $A$ and $B$ have the same Jordan form $J$, then $A\sim J\sim B$. The rank condition determines the Jordan form by (2). ∎
Over $\mathbb{R}$ (or $\mathbb{Q}$) a matrix whose characteristic polynomial does not split — a rotation, say — has no Jordan form over that field; it still has one over $\mathbb{C}$, and the rational canonical form below works over every field. When people say "the Jordan form of $A$" for a real matrix they mean the complex one.
- The number of Jordan blocks for $\lambda$ is the geometric multiplicity $g_\lambda=n-\operatorname{rank}(A-\lambda I)$.
- The total size of the blocks for $\lambda$ is the algebraic multiplicity $a_\lambda=\dim G_\lambda(A)$.
- The size of the largest block for $\lambda$ is the exponent $s_\lambda$ of $(x-\lambda)$ in $m_A$, which is also the least $s$ with $\operatorname{rank}(A-\lambda I)^s=n-a_\lambda$.
- $p_A=\prod_\lambda(x-\lambda)^{a_\lambda}$ and $m_A=\prod_\lambda(x-\lambda)^{s_\lambda}$; and $A$ is diagonalisable $\iff$ all blocks are $1\times1$ $\iff$ $m_A$ has no repeated root $\iff$ $g_\lambda=a_\lambda$ for every $\lambda$.
Proof. (1) is Theorem 9.25(2) with $k=1$: the number of blocks for $\lambda$ is $r_0-r_1=n-\operatorname{rank}(A-\lambda I)$. (2) The chains for $\lambda$ form a basis of $G_\lambda$ (they lie in it, are independent, and the chains for the other eigenvalues lie in the other $G_\mu$), and $\dim G_\lambda=a_\lambda$ by Theorem 9.22(2). (3) On $G_\lambda$, $N=A-\lambda I$ has index equal to the largest block (Theorem 9.24(3)), which equals $s_\lambda$ by Theorem 9.21; and $\operatorname{rank}(A-\lambda I)^s$ decreases until it reaches its stable value $\dim\operatorname{im}N^{s_\lambda}=n-a_\lambda$ (Fitting decomposition). (4) The formula for $p_A$ is Theorem 9.23(1) applied block by block, the one for $m_A$ is (3), and $A$ is diagonalisable iff it is similar to a Jordan matrix with $1\times1$ blocks only, iff every $s_\lambda=1$, iff (by (1), (2)) every $g_\lambda=a_\lambda$. This agrees with the criterion of §9.2. ∎
- Eigenvalues. Find $p_A$, its roots $\lambda$ and the multiplicities $a_\lambda$. (Shortcuts: trace, determinant, triangular structure; if some power of $A-\lambda I$ is $0$, then $\lambda$ is the only eigenvalue.)
- Ranks. For each $\lambda$, compute $r_j=\operatorname{rank}(A-\lambda I)^j$ for $j=1,2,\dots$ until $r_j=n-a_\lambda$. Then the $j$-th column of the dot diagram for $\lambda$ has $r_{j-1}-r_j$ dots. Check: the columns add up to $a_\lambda$ and the jumps $r_{j-1}-r_j$ never increase.
- Read off the blocks (rows of the dot diagram), $g_\lambda$ (first column) and $m_A$ (length of the first row).
- Jordan basis, if asked. Work top-down: for the longest chains choose tops $v\in\ker N^{s}\setminus\ker N^{s-1}$, independent modulo $\ker N^{s-1}$, and list $N^{s-1}v,\dots,Nv,v$. For shorter chains choose tops in $\ker N^{j}$ independent of $\ker N^{j-1}$ and of the vectors already used. In small examples: the bottoms must form a basis of $\ker N$ together with the other bottoms.
- Assemble $P$ with each chain as consecutive columns, bottom first; then $P^{-1}AP=J$ with ones above the diagonal. Check $AP=PJ$ column by column: $Av_1=\lambda v_1$, $Av_j=\lambda v_j+v_{j-1}$.
- If every eigenvalue has algebraic multiplicity at most $3$ (in particular if $n\le3$), then $A\sim B$.
- If in addition $g_\lambda(A)=g_\lambda(B)$ for every $\lambda$, and every eigenvalue has algebraic multiplicity at most $6$ (in particular if $n\le6$), then $A\sim B$.
- Both bounds are sharp: $(2,2)$ versus $(2,1,1)$ for (1), and $(3,3,1)$ versus $(3,2,2)$ for (2).
Proof. By Theorem 9.25 we must show that, for each eigenvalue, the partition of $a=a_\lambda$ formed by the block sizes is determined by its largest part $m$ (the exponent in $m_A$) — and in (2) also by its number of parts $t=g_\lambda$. (1) was checked in Example 4(c). (2) Two partitions with the same largest part $m$ and the same number $t$ of parts differ only in their remaining $t-1$ parts, which lie in $\{1,\dots,m\}$ and add up to $a-m$. Write these parts as $1+e_i$ with $0\le e_i\le m-1$ and $\sum e_i=e:=a-m-(t-1)$. If every $e_i\le1$ in both partitions, each is determined by $e$ (it has $e$ parts equal to $2$ and the rest equal to $1$); so two different partitions need some $e_i\ge2$ (forcing $m\ge3$), $e\ge2$, and at least two remaining parts ($t-1\ge2$) to redistribute. Then $a=m+(t-1)+e\ge3+2+2=7$. So for $a\le6$ the partition is determined. (3) was shown in Example 4(a), (b). ∎
Counting Similarity Classes and Using the Jordan Form
Proof. By Theorem 9.25, similarity classes correspond to Jordan matrices up to the order of blocks, i.e. to a choice, for each $\lambda_i$, of a partition of $a_i$ (the block sizes). The condition on $m_A$ says that this partition has largest part $m_i$ (Theorem 9.26(3)); the choices for different eigenvalues are independent, so we multiply. The nilpotent statements are Theorem 9.24(4), with $N^k=0$ meaning that all blocks have size at most $k$. ∎
| $a$ | $P(a,1)$ | $P(a,2)$ | $P(a,3)$ | $P(a,4)$ | $P(a,5)$ | $P(a,6)$ | $P(a,7)$ | $p(a)$ |
|---|---|---|---|---|---|---|---|---|
| 1 | 1 | 1 | ||||||
| 2 | 1 | 1 | 2 | |||||
| 3 | 1 | 1 | 1 | 3 | ||||
| 4 | 1 | 2 | 1 | 1 | 5 | |||
| 5 | 1 | 2 | 2 | 1 | 1 | 7 | ||
| 6 | 1 | 3 | 3 | 2 | 1 | 1 | 11 | |
| 7 | 1 | 3 | 4 | 3 | 2 | 1 | 1 | 15 |
Partitions of $a$ by largest part. Useful identities: $P(a,1)=P(a,a)=1$, $P(a,a-1)=1$ for $a\ge2$, $P(a,2)=\lfloor a/2\rfloor$, and $P(a,m)$ equals the number of partitions of $a-m$ into parts of size at most $m$ (remove the largest part).
- $A$ is similar to its transpose $A^{T}$.
- Jordan–Chevalley decomposition. There are unique $D,N\in M_n(\mathbb{C})$ with $A=D+N$, $D$ diagonalisable, $N$ nilpotent and $DN=ND$. Moreover $D$ and $N$ are polynomials in $A$, and $D$ has the same eigenvalues (with multiplicities) as $A$.
- Powers. $A^m\to0$ as $m\to\infty$ iff every eigenvalue satisfies $|\lambda|\lt1$. The sequence $(A^m)$ is bounded iff every $|\lambda|\le1$ and every Jordan block for an eigenvalue with $|\lambda|=1$ is $1\times1$.
Proof. (1) Write $A=PJP^{-1}$ with $J$ a Jordan matrix. Let $R$ be block diagonal with the reversal matrix of each block's size; Theorem 9.23(5) gives $RJR=J^{T}$ with $R=R^{-1}$. So $A^{T}=(P^{-1})^{T}J^{T}P^{T}\sim J^{T}\sim J\sim A$.
(2) Existence. Split $J=\Delta+M$ with $\Delta$ the diagonal part and $M$ the superdiagonal part. In each block $\Delta=\lambda I$ commutes with $M=J_k(0)$, so $\Delta M=M\Delta$; put $D=P\Delta P^{-1}$, $N=PMP^{-1}$. Polynomials. By the primary decomposition (§9.2) the projection $E_i$ of $V$ onto $G_{\lambda_i}$ along the other generalised eigenspaces is a polynomial in $A$. Since $D$ acts as $\lambda_i$ on $G_{\lambda_i}$, $D=\sum_i\lambda_iE_i$ is a polynomial in $A$, and so is $N=A-D$. Uniqueness. Let $A=D'+N'$ be another such decomposition. Then $D'$ commutes with $A=D'+N'$, hence with every polynomial in $A$, in particular with $D$ and $N$; similarly for $N'$. Now $D-D'=N'-N$. The left side is diagonalisable, since commuting diagonalisable matrices are simultaneously diagonalisable (§9.2); the right side is nilpotent, since a sum of commuting nilpotents is nilpotent (§8.2). A diagonalisable nilpotent matrix is similar to a diagonal matrix with zero diagonal, i.e. it is $0$. So $D=D'$ and $N=N'$.
(3) $A^m=PJ^mP^{-1}$, so it suffices to look at one block $J_k(\lambda)$, whose $m$-th power has entries $\binom{m}{j}\lambda^{m-j}$, $0\le j\le k-1$ (Theorem 9.23(4)). If $|\lambda|\lt1$, then $\binom{m}{j}|\lambda|^{m-j}\le m^j|\lambda|^{m-j}\to0$, because $m^j|\lambda|^m\to0$ (the standard limit $n^ka^n\to0$ for $|a|\lt1$, §2.1). If $|\lambda|\ge1$ the diagonal entries $\lambda^m$ do not tend to $0$. For boundedness: if $|\lambda|\gt1$ the diagonal is unbounded; if $|\lambda|=1$ and $k\ge2$ the entry $m\lambda^{m-1}$ has modulus $m\to\infty$; if $|\lambda|=1$ and $k=1$ the block is $(\lambda^m)$, of modulus $1$; and blocks with $|\lambda|\lt1$ tend to $0$. ∎
Part (3) hides a phenomenon that matters in practice: with $|\lambda|\lt1$ the powers do tend to $0$, but the off-diagonal entry $m\lambda^{m-1}$ can first become large — for $\lambda=0.9$ it peaks near $3.87$ at $m=9,10$. Interactive 2 shows this "transient growth" and the drift along the single eigenvector.
The Rational Canonical Form (in brief)
The Jordan form needs the eigenvalues to lie in $\mathbb{F}$. The rational canonical form replaces Jordan blocks by companion matrices, which exist over every field, and it decides similarity over $\mathbb{Q}$ or $\mathbb{R}$ without ever factorising a polynomial.
For $v\in V$, the $T$-cyclic subspace generated by $v$ is $Z(v;T)=\operatorname{span}\{v,Tv,T^2v,\dots\}$, and the $T$-annihilator of $v$ is the monic polynomial $f$ of least degree with $f(T)v=0$ (it divides $m_T$). The companion matrix of a monic $f(x)=x^k+a_{k-1}x^{k-1}+\cdots+a_1x+a_0$ is $$C(f)=\begin{pmatrix}0&0&\cdots&0&-a_0\\1&0&\cdots&0&-a_1\\0&1&\cdots&0&-a_2\\ \vdots&&\ddots&&\vdots\\0&0&\cdots&1&-a_{k-1}\end{pmatrix}$$ (some books use its transpose, which is similar to it). A matrix is in rational canonical form if it is $\operatorname{diag}\big(C(f_1),\dots,C(f_r)\big)$ with monic nonconstant $f_1\mid f_2\mid\cdots\mid f_r$; the $f_i$ are its invariant factors.
- If $v$ has $T$-annihilator $f$ of degree $k$, then $v,Tv,\dots,T^{k-1}v$ is a basis of $Z(v;T)$, in which $T|_{Z(v;T)}$ has matrix $C(f)$; and $p_{C(f)}=m_{C(f)}=f$.
- Existence and uniqueness (over any field). $V=Z(v_1;T)\oplus\cdots\oplus Z(v_r;T)$ for suitable $v_i$ whose annihilators satisfy $f_1\mid f_2\mid\cdots\mid f_r$; the $f_i$ are uniquely determined by $T$, and $f_r=m_T$, $f_1f_2\cdots f_r=p_T$. Hence every $A\in M_n(\mathbb{F})$ is similar over $\mathbb{F}$ to exactly one matrix in rational canonical form.
- Computing them. Let $\Delta_k(x)$ be the monic gcd of all $k\times k$ minors of $xI-A$ ($\Delta_0=1$). Then the invariant factors are the nonconstant ones among $\Delta_k/\Delta_{k-1}$, $k=1,\dots,n$. Consequently, if $A,B\in M_n(\mathbb{F})$ are similar over a larger field $\mathbb{K}\supseteq\mathbb{F}$, they are similar over $\mathbb{F}$; and $A\sim A^{T}$ over every field.
- From the Jordan form (when $p_A$ splits). For each eigenvalue list its block sizes $k_1(\lambda)\ge k_2(\lambda)\ge\cdots$ (padding with zeros). Then $$f_r=\prod_\lambda(x-\lambda)^{k_1(\lambda)},\quad f_{r-1}=\prod_\lambda(x-\lambda)^{k_2(\lambda)},\ \dots,$$ and $r=\max_\lambda g_\lambda$. The factors $(x-\lambda)^{k}$, one for each Jordan block, are the elementary divisors of $A$.
Proof. (1) If $\sum_{i\lt k}c_iT^iv=0$ with some $c_i\ne0$, a nonzero polynomial of degree $\lt k$ would kill $v$; so $v,\dots,T^{k-1}v$ are independent. Since $T^kv=-\sum_{i\lt k}a_iT^iv$ lies in their span, the span is $T$-invariant and therefore equals $Z(v;T)$. The matrix: $T$ sends the $i$-th basis vector $T^{i-1}v$ to the next one, and the last one to $-\sum a_iT^iv$ — that is $C(f)$. For $C=C(f)$ itself: $e_1,Ce_1=e_2,\dots,C^{k-1}e_1=e_k$ are independent, so no nonzero polynomial of degree $\lt k$ kills $C$ and $\deg m_C\ge k$; and $f(C)e_1=0$ (by the shape of the last column), whence $f(C)C^je_1=C^jf(C)e_1=0$ for all $j$, so $f(C)=0$. Thus $m_C=f$, and $p_C$, monic of degree $k$ and divisible by $m_C$ (Cayley–Hamilton), equals $f$.
(2), (3) Sketch. Make $V$ a module over the polynomial ring $\mathbb{F}[x]$ by $x\cdot v=Tv$. It is finitely generated and torsion, so by the structure theorem for finitely generated modules over a principal ideal domain (principal ideal domains such as $\mathbb{F}[x]$ are studied in Unit 2) it is a direct sum of cyclic modules $\mathbb{F}[x]/(f_i)$ with $f_1\mid\cdots\mid f_r$, uniquely; cyclic submodules are exactly the cyclic subspaces, and (1) turns the decomposition into the block form. Computationally, row and column operations over $\mathbb{F}[x]$ bring $xI-A$ to its Smith normal form $\operatorname{diag}(1,\dots,1,f_1,\dots,f_r)$, and such operations do not change the gcds $\Delta_k$ of the $k\times k$ minors. These gcds are computed by polynomial arithmetic in $\mathbb{F}[x]$, so they are the same whether we regard $A$ as a matrix over $\mathbb{F}$ or over $\mathbb{K}$ — that is why similarity does not depend on the field. And $xI-A^{T}=(xI-A)^{T}$ has the same minors as $xI-A$. The last two relations in (2) follow from (1): $m_T$ is the lcm of the $m_{C(f_i)}=f_i$, which is $f_r$, and $p_T=\prod p_{C(f_i)}=\prod f_i$.
(4) We use a lemma: if $u,w$ have coprime $T$-annihilators $g,h$, then $u+w$ has annihilator $gh$ and $Z(u+w;T)=Z(u;T)\oplus Z(w;T)$. Indeed, $Z(u)\cap Z(w)$ is killed by $g$ and by $h$, hence by $\gcd(g,h)=1$, so it is $\{0\}$; if $f(T)(u+w)=0$ then $f(T)u=-f(T)w\in Z(u)\cap Z(w)=\{0\}$, so $g\mid f$ and $h\mid f$, i.e. $gh\mid f$; conversely $gh$ kills $u+w$. Then $\dim Z(u+w)=\deg gh$ $=\dim Z(u)+\dim Z(w)$ while $Z(u+w)\subseteq Z(u)\oplus Z(w)$, so they are equal. Now, for each $j$, take the top $w_{j,\lambda}$ of the $j$-th longest chain for each eigenvalue $\lambda$ that has one; its annihilator is $(x-\lambda)^{k_j(\lambda)}$ and its cyclic subspace is that chain. By the lemma $v_j=\sum_\lambda w_{j,\lambda}$ has annihilator $\prod_\lambda(x-\lambda)^{k_j(\lambda)}$ and $Z(v_j)$ is the sum of those chains. These subspaces fill $V$, and the annihilator of $v_{j+1}$ divides that of $v_j$ because $k_{j+1}(\lambda)\le k_j(\lambda)$ for every $\lambda$. Listing the $v_j$ in reverse order gives $f_1\mid\cdots\mid f_r$, and by uniqueness in (2) these are the invariant factors. ∎
- Same $p_A$ and $m_A$ do not give similarity once $a_\lambda\ge4$: $J_2(0)\oplus J_2(0)$ and $J_2(0)\oplus J_1(0)\oplus J_1(0)$ ($p=x^4$, $m=x^2$, ranks $2$ and $1$).
- Adding the geometric multiplicity still fails from $7$ on: types $(3,3,1)$ and $(3,2,2)$ share $p$, $m$ and $g_0=3$; only $\operatorname{rank}N^2$ ($2$ versus $1$) separates them.
- Which invariant counts what: $g_\lambda$ = number of blocks; $a_\lambda$ = total size; exponent in $m_A$ = size of the largest block. Mixing these up is the commonest error.
- One rank is not enough: $\operatorname{rank}(A-\lambda I)$ gives only the number of blocks; the sizes need the whole sequence $\operatorname{rank}(A-\lambda I)^k$.
- The ranks stabilise at $n-a_\lambda$, not at $0$, when $A$ has other eigenvalues (Example 3: $\operatorname{rank}(A-3I)^k=3$ for $k\ge3$).
- Jordan forms need split characteristic polynomials: the rotation $\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ has no Jordan form over $\mathbb{R}$ (its form over $\mathbb{C}$ is $\operatorname{diag}(i,-i)$); use the rational canonical form over $\mathbb{R}$.
- Order the chain bottom first: columns $(N^{k-1}v,\dots,Nv,v)$ give ones above the diagonal; the reverse order gives $J^{T}$.
- Not every choice of chain vectors works: the bottoms of all chains must be independent. In Example 2, using $(1,1,0,0)=-N^2e_1$ as the second chain makes $P$ singular.
- Squares of nilpotent blocks split: $J_k(0)^2$ has blocks of sizes $\lceil k/2\rceil$ and $\lfloor k/2\rfloor$, whereas $J_k(\lambda)^2\sim J_k(\lambda^2)$ for $\lambda\ne0$.
- $A^2$ diagonalisable does not make $A$ diagonalisable ($A=J_2(0)$), unless $A$ is invertible (Example 6(c)).
- $|\lambda|\lt1$ does not mean $A^m$ shrinks at once: $J_2(0.9)^{10}$ has corner entry $10\cdot0.9^9\approx3.87$.
- $A\sim A^{T}$ always, but $A\not\sim A^{*}$ in general: $A=(i)$ has $A^{*}=(-i)$ — conjugation changes the eigenvalues.
- The Jordan form is not continuous: $J_n(0)+\varepsilon E_{n1}$ has the $n$ distinct eigenvalues $\varepsilon^{1/n}e^{2\pi ik/n}$, so it is diagonalisable for every $\varepsilon\ne0$. Numerical software therefore avoids computing Jordan forms (see the Research–Teaching Nexus).
Exercises
- 1 For $J=J_4(5)$ compute $\operatorname{rank}(J-5I)^j$ for $j=0,1,2,3,4$, and describe all generalised eigenvectors of height exactly $2$.
- 2 Write out $J_3(\lambda)^m$ for $m\ge2$ and compute $J_3(2)^{10}$.
- 3 For $A=\begin{pmatrix}1&-2&-2\\-1&2&1\\1&1&2\end{pmatrix}$ find $p_A$, the generalised eigenspaces, the stabilisation indices and $m_A$, and verify the Fitting decomposition for $\lambda=1$.
- 4 Let $N=T-\lambda I$ and let $v$ have height $k$. Prove that $v,Nv,\dots,N^{k-1}v$ are linearly independent, that they span a $T$-invariant subspace $Z$, and that $T|_Z$ has matrix $J_k(\lambda)$ in the basis $(N^{k-1}v,\dots,Nv,v)$.
- 5 Let $S\in\mathcal{L}(V)$ commute with $T$. Prove that $S$ maps $G_\lambda(T)$ and $\operatorname{im}(T-\lambda I)^n$ into themselves.
- 6 Show that for $\lambda\ne0$, $J_k(\lambda)^{-1}\sim J_k(1/\lambda)$.
- 7 Find the Jordan form of $D=d/dx$ on $P_n(\mathbb{R})$, and on $W=\operatorname{span}\{e^x,xe^x,x^2e^x\}$ (give a Jordan basis of $W$).
- 8 For an eigenvalue $\lambda$ of $A\in M_n(\mathbb{C})$, prove that the following are equivalent: (i) all Jordan blocks for $\lambda$ are $1\times1$; (ii) $\ker(A-\lambda I)=\ker(A-\lambda I)^2$; (iii) $\operatorname{rank}(A-\lambda I)=\operatorname{rank}(A-\lambda I)^2$.
- 1 Find the Jordan form $J$ of $A=\begin{pmatrix}3&1\\-1&1\end{pmatrix}$ and an invertible $P$ with $P^{-1}AP=J$.
- 2 Do the same for $A=\begin{pmatrix}1&-1&-1\\2&4&1\\2&1&4\end{pmatrix}$, and find $m_A$.
- 3 Do the same for $A=\begin{pmatrix}-1&-1&-1\\-1&-1&0\\1&0&-1\end{pmatrix}$.
- 4 Find the Jordan form and the minimal polynomial (no basis needed) of $A=\begin{pmatrix}-2&-3&3&0\\3&3&-2&0\\3&2&-1&0\\3&2&1&-2\end{pmatrix}$, given that $p_A=(x-1)^2(x+2)^2$.
- 5 Let $U\in M_n(\mathbb{R})$ have all entries above the diagonal equal to $1$ and all other entries $0$. Find its Jordan form.
- 6 Find the Jordan form of $T:P_3(\mathbb{R})\to P_3(\mathbb{R})$, $T(p)(x)=p(x+1)$.
- 7 Find the Jordan form of $L_A:M_2(\mathbb{C})\to M_2(\mathbb{C})$, $L_A(X)=AX$, for $A=J_2(0)$.
- 8 $A\in M_6(\mathbb{C})$ has $0$ as its only eigenvalue, $\operatorname{rank}A=3$, $\operatorname{rank}A^2=1$ and $A^3=0$. Find its Jordan form and $m_A$.
- 9 Use the Jordan form of II.1 to compute $A^{100}$.
- 1 List all possible Jordan forms of $A\in M_5(\mathbb{C})$ with $p_A=(x-2)^3(x+1)^2$ and $m_A=(x-2)^2(x+1)$.
- 2 How many similarity classes in $M_6(\mathbb{C})$ have $p_A=(x-1)^6$ and $m_A=(x-1)^3$? List them.
- 3 How many similarity classes of nilpotent $5\times5$ complex matrices are there? How many satisfy $N^2=0$?
- 4 How many similarity classes of $A\in M_4(\mathbb{C})$ satisfy $A^2=A^3$?
- 5 Prove directly from the Jordan form that two $3\times3$ complex matrices are similar iff they have the same characteristic and minimal polynomials, and explain why the "only if" direction holds for all $n$.
- 6 $A\in M_7(\mathbb{C})$ has $p_A=x^7$, $\operatorname{rank}A=4$ and $\operatorname{rank}A^2=1$. Find its Jordan form and $m_A$. Would $\operatorname{rank}A=4$ and $m_A$ alone have sufficed?
- 7 Prove that $A^2=0$ implies $\operatorname{rank}A\le n/2$, and that there are exactly $\lfloor n/2\rfloor+1$ similarity classes of $n\times n$ complex matrices with $A^2=0$.
- 8 How many similarity classes of $A\in M_3(\mathbb{C})$ satisfy $A^3=I$?
- 1 Prove that every $A\in M_n(\mathbb{C})$ is similar to $A^{T}$, first for a single Jordan block and then in general.
- 2 Prove that if $A\in M_n(\mathbb{C})$ and $A^k=I$ for some $k\ge1$, then $A$ is diagonalisable. Is a real matrix with $A^k=I$ diagonalisable over $\mathbb{R}$?
- 3 Let $N\in M_n(\mathbb{F})$ be nilpotent with $N^{n-1}\ne0$. Show that $N\sim J_n(0)$, and deduce that every nilpotent $n\times n$ matrix of rank $n-1$ is similar to $J_n(0)$.
- 4 Let $N=J_k(0)$, $m\ge1$, and write $k=qm+r$ with $0\le r\lt m$. Show that $N^m$ has $r$ Jordan blocks of size $q+1$ and $m-r$ of size $q$ (omit blocks of size $0$). Find the Jordan forms of $J_7(0)^3$ and $J_6(0)^2$.
- 5 $A\in M_4(\mathbb{Q})$ has $p_A=(x^2-2)^2$. Find its rational canonical form over $\mathbb{Q}$ and its Jordan form over $\mathbb{R}$, in the two cases $m_A=x^2-2$ and $m_A=(x^2-2)^2$.
- 6 Find the invariant factors and the rational canonical form of $J_2(1)\oplus J_1(1)\oplus J_1(2)$.
- 7 Show that every $A\in M_n(\mathbb{C})$ is similar, for each $\varepsilon\gt0$, to a matrix that agrees with the Jordan form of $A$ except that the superdiagonal ones are replaced by $\varepsilon$.
- 8 Let $A\in M_n(\mathbb{F})$ with $m_A=p_A$. Show that there is $v$ with $v,Av,\dots,A^{n-1}v$ a basis, and that every $B$ commuting with $A$ is a polynomial in $A$.
E1. CSIR Part B GATE MA Let $A\in M_6(\mathbb{C})$ with characteristic polynomial $(x-2)^6$, $\operatorname{rank}(A-2I)=3$ and $\operatorname{rank}(A-2I)^2=1$. The number of Jordan blocks of size $2$ in the Jordan form of $A$ is (A) $0$; (B) $1$; (C) $2$; (D) $3$.
E2. CSIR Part B The number of similarity classes of $6\times6$ complex matrices $A$ with $A^3=0$ but $A^2\ne0$ is (A) $2$; (B) $3$; (C) $4$; (D) $7$.
E3. CSIR Part B Let $N\in M_5(\mathbb{R})$ be nilpotent of rank $3$. Which of the following is necessarily true? (A) $N^3=0$; (B) $N^4=0$; (C) $N^4\ne0$; (D) $N^2=0$.
E4. CSIR Part C Let $A\in M_4(\mathbb{C})$ have characteristic polynomial $(x-1)^2(x-3)^2$. Which of the following statements are true? (One or more options may be correct.) (A) $A$ is diagonalisable if and only if $(A-I)(A-3I)=0$. (B) If $\operatorname{rank}(A-I)=3$, then $\operatorname{rank}(A-3I)=3$. (C) There are exactly $4$ similarity classes of such matrices. (D) $\operatorname{rank}(A-I)^2=2$.
E5. CSIR Part C GATE MA Let $n\ge2$. Which of the following hold for every nilpotent $N\in M_n(\mathbb{C})$? (A) $N$ is similar to $3N$. (B) $I+N$ is similar to $(I+N)^{-1}$. (C) $N$ is similar to $N^2$. (D) $N+N^{T}$ is nilpotent.
E6. CSIR Part C Let $A\in M_4(\mathbb{R})$ have minimal polynomial $(x^2+1)^2$. Which of the following are true? (A) $A$ is diagonalisable over $\mathbb{C}$. (B) The characteristic polynomial of $A$ is $(x^2+1)^2$. (C) $\operatorname{rank}(A^2+I)=2$. (D) $A$ is similar over $\mathbb{R}$ to $\operatorname{diag}(B,B)$ for some $B\in M_2(\mathbb{R})$.
Suggested plan — 3 lectures × 60 min.
- Generalised eigenspaces. Start from $\begin{pmatrix}2&1\\0&2\end{pmatrix}$ and ask what $e_2$ "is". Definition 9.10, Theorem 9.21 (prove (1)–(3) fully; (4) can be done as "the kernels stop growing exactly when the factor $(x-\lambda)$ has been used up"), Theorem 9.22, Example 1 on the board.
- Nilpotent operators and the Jordan form. Theorem 9.23 by direct computation; Definition 9.12 with dot diagrams drawn for $(3,2,2)$ and $(3,3,1)$; the proof of Theorem 9.24 (draw $U=\operatorname{im}N$ as the dot diagram with the first column removed — the induction then "adds a column back"). Project Interactive 1, first with $(3,2,2)$ then $(3,3,1)$, and ask the class to predict $\operatorname{rank}N^2$ before stepping. Theorems 9.25–9.26, the strategy box and Example 2.
- Using it. Examples 3–5 as rapid-fire board work; Theorem 9.27 with the partition argument; Theorem 9.28 and the table. Theorem 9.29 with Interactive 2 ($\lambda=0.9$, then $1$, then $-0.9$, then switch to $\lambda I$). Close with the rational canonical form (Example 7) and Exam-corner E4, E6 done live.
Discussion questions. (1) Why is "number of blocks of size $\ge j$" so much easier to read from ranks than "number of blocks of size $j$"? (2) A matrix with real entries can have a Jordan form with non-real entries; can a real matrix be similar over $\mathbb{C}$ but not over $\mathbb{R}$ to another real matrix? (No — Theorem 9.30(3).) (3) Arbitrarily small perturbations make every matrix diagonalisable; is the Jordan form then "physically meaningful"? (Discuss Interactive 2's transient growth.)
Misconceptions to watch for. Confusing $g_\lambda$ (number of blocks) with the largest block; believing $p$ and $m$ always determine similarity; stopping the rank sequence at the first repeated value without checking it equals $n-a_\lambda$; putting chain vectors in $P$ in the wrong order; thinking $J_k(0)^2$ is again one block; assuming that a real matrix always has a real Jordan form.
Suggested homework. I(1, 3, 6, 7), II(2, 3, 4, 8, 9), III(1, 2, 4, 6, 8), IV(1, 4, 6); Exam corner E1, E3, E5.
Research question. How fragile is a Jordan block? If a defective matrix is perturbed by an amount $\varepsilon$, how far do its eigenvalues move — and why do numerical libraries compute Schur forms rather than Jordan forms?
Why it matters. Engineers meet Jordan blocks whenever two modes of a system coalesce: the flutter boundary of an aircraft wing, where two vibration frequencies merge; "exceptional points" in lasers and optical resonators, which physicists exploit because the eigenvalues there respond to tiny disturbances like $\sqrt\varepsilon$ rather than $\varepsilon$; and critically damped feedback loops in control engineering. The same sensitivity is why standard numerical software computes eigenvalues through the (stable) Schur triangularisation of §9.2 and treats the Jordan form as a symbolic, exact-arithmetic object.
What to do.
- Show that $J_n(0)+\varepsilon E_{n1}$ (put $\varepsilon$ in the bottom-left corner) is the companion matrix of $x^n-\varepsilon$, so its eigenvalues are the $n$ complex $n$-th roots of $\varepsilon$, all of modulus $\varepsilon^{1/n}$.
- With
numpy.linalg.eigvals, compute these eigenvalues for $n=4,8,16$ and $\varepsilon=10^{-8},10^{-12}$. Tabulate the mean modulus against $\varepsilon^{1/n}$ and plot the eigenvalues in the complex plane. - Repeat with a random perturbation $\varepsilon R$ (entries of $R$ uniform in $[-1,1]$) of the diagonal matrix $\operatorname{diag}(1,2,\dots,n)$: how far do the eigenvalues move now?
- Ask the computer for the eigenvalues of the matrix $A$ of Example 2 (exactly $2,2,2,2$) in double precision and measure the error. Explain the size of the error with step 1, using the block $J_3(2)$ and machine precision $\approx10^{-16}$.
- Use
sympy's exactjordan_formon $A$, then on $A$ with one entry changed by $10^{-12}$ (as an exact rational). Compare the two answers.
Tools: Python with numpy, sympy and matplotlib (or any computer algebra system).
Hand in: A one-page note with the table of step 2, the error of step 4 with your explanation, and two sentences on what "the Jordan form of a measured matrix" can mean.
Going further. Read about pseudospectra (the sets where $\|(zI-A)^{-1}\|$ is large) and the Bauer–Fike theorem, which bounds eigenvalue perturbations for diagonalisable matrices by the condition number of the eigenvector matrix — and explain why that bound must blow up as a matrix approaches a Jordan block.
🧪 Self-check quiz
Q1. For an eigenvalue $\lambda$ of $A\in M_n(\mathbb{C})$, the number of Jordan blocks for $\lambda$ equals
Q2. The size of the largest Jordan block for $\lambda$ equals
Q3. How many similarity classes of nilpotent $4\times4$ complex matrices are there?
Q4. $A\in M_5(\mathbb{C})$ has $p_A=(x-1)^5$, $\operatorname{rank}(A-I)=2$ and $(A-I)^2=0$. Its Jordan form is
Q5. Which pair of matrices is similar?