Chapter 9 · Eigenvalues and Canonical Forms

9.2 Diagonal and Triangular Forms

A diagonalisable operator is a list of independent stretchings. The minimal polynomial decides exactly when that happens; when it does not, the primary decomposition still splits the space into pieces with one eigenvalue each, and every operator whose characteristic polynomial splits is triangular in a suitable — even orthonormal — basis.

If $A=PDP^{-1}$ with $D=\operatorname{diag}(\lambda_1,\dots,\lambda_n)$, then every computation with $A$ becomes a computation with numbers: $A^k=PD^kP^{-1}$, $f(A)=Pf(D)P^{-1}$, $e^{tA}=Pe^{tD}P^{-1}$, and the system $x'=Ax$ splits into $n$ independent equations $y_i'=\lambda_iy_i$ in the coordinates $y=P^{-1}x$. Engineers use this to separate the vibration modes of a structure, chemists to decouple reaction rates, statisticians to compute long-run behaviour of Markov chains. The trouble is that not every matrix can be diagonalised — the shear $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ has only one eigen-direction — and that whether a matrix can be diagonalised may depend on the field.

This section answers the question completely. A matrix is diagonalisable exactly when its minimal polynomial is a product of distinct linear factors (Theorem 9.16); the proof goes through the primary decomposition theorem, which splits $V$ into invariant pieces, one for each irreducible factor of $m_T$. When diagonalisation fails, triangularisation still works as soon as the characteristic polynomial splits — over $\mathbb{C}$ always, and by Schur's theorem even with a unitary change of basis. Finally, commuting operators can be diagonalised or triangularised simultaneously. CSIR NET asks constantly "which of these matrices are diagonalisable (over $\mathbb{R}$? over $\mathbb{C}$?)", "if $A^k=I$ then …", "if $AB=BA$ then …" — and this section is where those questions are settled. We use the notation of §9.1: $p_T$, $m_T$, $E_\lambda$, $a(\lambda)$, $g(\lambda)$.

Diagonalisable Operators

📘 Definition 9.6 — Diagonalisable operator and matrix
$T\in\mathcal{L}(V)$ is diagonalisable if $V$ has an ordered basis $\mathcal{B}$ for which $[T]_{\mathcal{B}}$ is diagonal. $A\in M_n(\mathbb{F})$ is diagonalisable over $\mathbb{F}$ if there is an invertible $P\in M_n(\mathbb{F})$ with $P^{-1}AP$ diagonal. The matrix $P$ diagonalises $A$.
📐 Theorem 9.12 — Five equivalent conditions for diagonalisability
Let $\lambda_1,\dots,\lambda_k$ be the distinct eigenvalues of $T$ in $\mathbb{F}$ ($\dim V=n$). The following are equivalent:
  1. $T$ is diagonalisable;
  2. $V$ has a basis consisting of eigenvectors of $T$;
  3. $V=E_{\lambda_1}\oplus\cdots\oplus E_{\lambda_k}$;
  4. $g(\lambda_1)+\cdots+g(\lambda_k)=n$;
  5. $p_T$ splits over $\mathbb{F}$ and $g(\lambda_i)=a(\lambda_i)$ for every $i$.
In that case, $P^{-1}AP=D$ exactly when the columns of $P$ form a basis of eigenvectors and the diagonal entries of $D$ are the corresponding eigenvalues; each $\lambda_i$ appears $a(\lambda_i)$ times on the diagonal.

Proof. (1)$\iff$(2): $[T]_{\mathcal{B}}=\operatorname{diag}(d_1,\dots,d_n)$ says precisely $Tb_j=d_jb_j$ for every basis vector $b_j$. (2)$\Rightarrow$(3): the sum of the eigenspaces is direct by Theorem 9.3, and it contains a basis, so it is $V$. (3)$\Rightarrow$(4): dimensions of a direct sum add up. (4)$\Rightarrow$(2): choose a basis of each $E_{\lambda_i}$; the union has $\sum g(\lambda_i)=n$ vectors, and it is independent because the sum of the eigenspaces is direct (Theorem 9.3); so it is a basis of eigenvectors. (4)$\iff$(5): by Theorem 9.4, $\sum g(\lambda_i)\le\sum a(\lambda_i)\le n$ with $g(\lambda_i)\le a(\lambda_i)$ termwise. So $\sum g=n$ holds iff $\sum a=n$ (i.e. $p_T$ splits) and every inequality $g(\lambda_i)\le a(\lambda_i)$ is an equality. For the last statement: $AP=PD$ says column by column that $Ap_j=d_jp_j$, and $P$ is invertible iff its columns form a basis; $p_D=\prod(x-d_j)$ equals $p_A$, so each $\lambda_i$ appears $a(\lambda_i)$ times. ∎

📐 Theorem 9.13 — Distinct eigenvalues suffice
If $T$ has $n=\dim V$ distinct eigenvalues in $\mathbb{F}$, then $T$ is diagonalisable. The condition is sufficient but not necessary ($I_n$ is diagonal with a single eigenvalue).

Proof. Each of the $n$ eigenvalues has $g\ge1$, so $\sum g\ge n$; with Theorem 9.4, $\sum g=n$, and Theorem 9.12(4) applies. ∎

✏️ Example 1 — Diagonalise, then compute every power
Diagonalise $D=\begin{pmatrix}3&-2&2\\2&-1&2\\2&-2&3\end{pmatrix}$ and find a closed formula for $D^k$.
Eigenvalue $1$ $D-I=\begin{pmatrix}2&-2&2\\2&-2&2\\2&-2&2\end{pmatrix}$ has rank $1$, so $g(1)=2$, and $E_1=\{v_1-v_2+v_3=0\}$ has basis $(1,1,0)$, $(-1,0,1)$.
Eigenvalue $3$ $\operatorname{tr}D=5=1+1+\lambda_3$ gives $\lambda_3=3$; $(D-3I)v=0$ gives $v=(1,1,1)$. Now $g(1)+g(3)=3$: $D$ is diagonalisable (Theorem 9.12(4)).
$P$ and $D$ $P=\begin{pmatrix}1&-1&1\\1&0&1\\0&1&1\end{pmatrix}$ (eigenvectors as columns), $\det P=1$, $P^{-1}=\begin{pmatrix}-1&2&-1\\-1&1&0\\1&-1&1\end{pmatrix}$, and $P^{-1}DP=\operatorname{diag}(1,1,3)$ — check by multiplying.
$D^k$ the quick way $D^k=P\operatorname{diag}(1,1,3^k)P^{-1}$. Faster: the minimal polynomial is $(x-1)(x-3)$ (we shall see why in Theorem 9.16), so $D^k=\alpha D+\beta I$ with $\alpha+\beta=1$, $3\alpha+\beta=3^k$, i.e. $$D^k=I+\tfrac{3^k-1}{2}\,(D-I)=I+(3^k-1)\begin{pmatrix}1&-1&1\\1&-1&1\\1&-1&1\end{pmatrix}.$$ Check $k=1$: $I+2\cdot\tfrac12(D-I)=D$. ✓
✏️ Example 2 — The field matters
Decide whether each matrix is diagonalisable over the given field: (a) $R=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ over $\mathbb{R}$ and over $\mathbb{C}$; (b) $N=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ over any field; (c) $S=\begin{pmatrix}0&1\\1&0\end{pmatrix}$ over $\mathbb{R}$ and over $\mathbb{F}_2$.
(a) $p_R=x^2+1$ does not split over $\mathbb{R}$: not diagonalisable over $\mathbb{R}$ (Theorem 9.12(5)). Over $\mathbb{C}$ it has the distinct roots $\pm i$, so it is diagonalisable: $R(1,-i)^T=(i,1)^T=i(1,-i)^T$, $R(1,i)^T=-i(1,i)^T$, and with $P=\begin{pmatrix}1&1\\-i&i\end{pmatrix}$, $P^{-1}RP=\operatorname{diag}(i,-i)$.
(b) $p_N=x^2$, $a(0)=2$, but $\operatorname{rank}N=1$ so $g(0)=1$: never diagonalisable. (A diagonalisable matrix with only the eigenvalue $0$ is similar to the zero matrix, hence is $0$.)
(c) Over $\mathbb{R}$ $p_S=x^2-1=(x-1)(x+1)$, distinct roots: diagonalisable, with eigenvectors $(1,1)$ and $(1,-1)$.
(c) Over $\mathbb{F}_2$ Now $-1=1$, so $p_S=x^2+1=(x+1)^2$ and the only eigenvalue is $1$. $S-I=\begin{pmatrix}1&1\\1&1\end{pmatrix}$ has rank $1$ over $\mathbb{F}_2$, so $g(1)=1\lt2$: not diagonalisable. The "eigenvectors" $(1,1)$ and $(1,-1)$ have become equal. Moral: the same integer matrix can be diagonalisable over one field and not over another.
Interactive 1 — When Eigenvectors Collide

The matrix $A_s=\begin{pmatrix}1&1\\s&1\end{pmatrix}$ has eigenvalues $1\pm\sqrt s$. On the left are its eigen-directions (green and purple) and the parallelogram spanned by the two unit eigenvectors — its area is $|\det P|$. On the right the two eigenvalues move in the complex plane. Press Play (or drag $s$): as $s\to0$ the eigenvectors merge, $\det P\to0$, and at $s=0$ the matrix is the shear — not diagonalisable. For $s\lt0$ the eigenvalues leave the real axis: diagonalisable over $\mathbb{C}$ but not over $\mathbb{R}$.

0.50
eigenvalues—
m_A(x) = p_A(x)—
|det P| (unit eigenvectors)—
diagonalisable over ℝ?—
over ℂ?—

Invariant Subspaces and the Primary Decomposition

📘 Definition 9.7 — Invariant subspace, restriction, induced operator
A subspace $W\subseteq V$ is $T$-invariant if $T(W)\subseteq W$. Then $T$ restricts to an operator $T|_W\in\mathcal{L}(W)$, and induces an operator $\overline{T}$ on the quotient $V/W$ by $\overline{T}(v+W)=Tv+W$ (well defined because $v-v'\in W$ implies $Tv-Tv'\in W$).
📐 Theorem 9.14 — Invariant subspaces and block matrices
  1. Let $W$ be $T$-invariant, and extend a basis $(w_1,\dots,w_r)$ of $W$ to a basis $\mathcal{B}$ of $V$. Then $[T]_{\mathcal{B}}=\begin{pmatrix}A&B\\0&C\end{pmatrix}$ with $A=[T|_W]$ and $C$ the matrix of $\overline{T}$ in the induced basis of $V/W$. Hence $p_T=p_{T|_W}\,p_{\overline{T}}$, and $m_{T|_W}$, $m_{\overline{T}}$ both divide $m_T$.
  2. If $V=W_1\oplus\cdots\oplus W_k$ with every $W_i$ invariant, then in a basis adapted to the decomposition $[T]$ is block diagonal, $p_T=\prod p_{T|_{W_i}}$ and $m_T=\operatorname{lcm}_i\,m_{T|_{W_i}}$.
  3. If $S\in\mathcal{L}(V)$ commutes with $T$, then $\ker S$, $\operatorname{im}S$ and every eigenspace of $S$ are $T$-invariant. In particular $\ker f(T)$ and $\operatorname{im}f(T)$ are $T$-invariant for every $f\in\mathbb{F}[x]$.

Proof. (1) $Tw_j\in W$ is a combination of $w_1,\dots,w_r$ only, so the first $r$ columns vanish below row $r$; the top-left block is $[T|_W]$. For $j\gt r$, $Tb_j=\sum_{i\le r}(\cdot)w_i+\sum_{i\gt r}c_{ij}b_i$, and modulo $W$ this reads $\overline{T}(b_j+W)=\sum_{i\gt r}c_{ij}(b_i+W)$, so $C=[\overline{T}]$. The factorisation of $p_T$ is Theorem 9.2(5). If $f(T)=0$ then $f(T)$ vanishes on $W$ and on $V/W$, i.e. $f(T|_W)=0$ and $f(\overline{T})=0$; so both minimal polynomials divide $m_T$ (Theorem 9.6). (2) is Theorem 9.6 for block diagonal matrices. (3) If $Sv=0$ then $S(Tv)=T(Sv)=0$; if $v=Su$ then $Tv=S(Tu)\in\operatorname{im}S$; and $E_\mu(S)=\ker(S-\mu I)$ with $S-\mu I$ commuting with $T$. Polynomials in $T$ commute with $T$ (Theorem 9.5). ∎

📐 Theorem 9.15 — The primary decomposition theorem
Let $m_T=q_1^{e_1}\cdots q_k^{e_k}$, where $q_1,\dots,q_k$ are distinct monic irreducible polynomials over $\mathbb{F}$, and put $W_i=\ker q_i(T)^{e_i}$. Then
  1. $V=W_1\oplus W_2\oplus\cdots\oplus W_k$, and each $W_i$ is $T$-invariant;
  2. the minimal polynomial of $T_i=T|_{W_i}$ is $q_i^{e_i}$;
  3. the projection $E_i$ of $V$ onto $W_i$ along $\bigoplus_{j\ne i}W_j$ is a polynomial in $T$.
In the split case $q_i=x-\lambda_i$: $W_i=\ker(T-\lambda_iI)^{e_i}$, the generalised eigenspace of $\lambda_i$, has dimension $a(\lambda_i)$, and $T_i=\lambda_iI+N_i$ with $N_i$ nilpotent of index $e_i$.

Proof. Put $f_i=m_T/q_i^{e_i}=\prod_{j\ne i}q_j^{e_j}$. No irreducible polynomial divides all of $f_1,\dots,f_k$ ($q_i$ does not divide $f_i$), so $\gcd(f_1,\dots,f_k)=1$, and by Bézout's identity in $\mathbb{F}[x]$ there are $g_1,\dots,g_k$ with $$g_1f_1+g_2f_2+\cdots+g_kf_k=1.$$ Define $E_i=g_i(T)f_i(T)$. Then:

So $W_i=\operatorname{im}E_i$, and every $v=\sum E_iv$ is a sum of elements of the $W_i$. The sum is direct: if $w_1+\cdots+w_k=0$ with $w_j\in W_j$, apply $E_i$; since $E_iw_j=E_iE_jw_j=0$ for $j\ne i$ (as $w_j=E_jw_j$) and $E_iw_i=w_i$, we get $w_i=0$. Each $W_i$ is invariant by Theorem 9.14(3). This proves (1) and (3).

(2) $q_i(T_i)^{e_i}=0$ by definition of $W_i$, so $m_{T_i}=q_i^{d_i}$ with $d_i\le e_i$ ($q_i$ is irreducible). By Theorem 9.14(2), $m_T=\operatorname{lcm}(q_1^{d_1},\dots,q_k^{d_k})=\prod q_i^{d_i}$; comparing with $m_T=\prod q_i^{e_i}$ gives $d_i=e_i$. In the split case $m_{T_i}=(x-\lambda_i)^{e_i}$, so $N_i=T_i-\lambda_iI$ is nilpotent of index $e_i$ and $p_{T_i}=(x-\lambda_i)^{\dim W_i}$ (its only root is $\lambda_i$, Theorem 9.8). Since $p_T=\prod p_{T_i}$ (Theorem 9.14(2)), the exponent of $x-\lambda_i$ in $p_T$ is $\dim W_i$, i.e. $\dim W_i=a(\lambda_i)$. ∎

📐 Theorem 9.16 — The minimal polynomial criterion
$T$ is diagonalisable over $\mathbb{F}$ if and only if $m_T=(x-\lambda_1)(x-\lambda_2)\cdots(x-\lambda_k)$ with distinct $\lambda_1,\dots,\lambda_k\in\mathbb{F}$. Then $T=\lambda_1E_1+\cdots+\lambda_kE_k$, where the projections onto the eigenspaces are given by the Lagrange polynomials $$E_i=\prod_{j\ne i}\frac{T-\lambda_jI}{\lambda_i-\lambda_j}.$$ Consequently: (a) if some polynomial with distinct roots, all in $\mathbb{F}$, annihilates $T$, then $T$ is diagonalisable; (b) the restriction of a diagonalisable $T$ to a $T$-invariant subspace is diagonalisable; (c) a diagonalisable nilpotent operator is $0$.

Proof. ($\Rightarrow$) Let $\mathcal{B}$ be a basis of eigenvectors and $f=\prod(x-\lambda_i)$ over the distinct eigenvalues. For $b\in\mathcal{B}$ with $Tb=\lambda_jb$, the factors of $f(T)$ commute and $(T-\lambda_jI)b=0$, so $f(T)b=0$. Hence $f(T)=0$ and $m_T\mid f$; since every eigenvalue is a root of $m_T$ (Theorem 9.8), $m_T=f$. ($\Leftarrow$) In Theorem 9.15 all $e_i=1$ and $q_i=x-\lambda_i$, so $V=\bigoplus\ker(T-\lambda_iI)=\bigoplus E_{\lambda_i}$, which is Theorem 9.12(3). For the formula: on $E_{\lambda_l}$ the operator $\prod_{j\ne i}(T-\lambda_jI)/(\lambda_i-\lambda_j)$ acts as the scalar $\prod_{j\ne i}(\lambda_l-\lambda_j)/(\lambda_i-\lambda_j)$, which is $1$ for $l=i$ and $0$ for $l\ne i$; that is the projection onto $E_{\lambda_i}$ along the others, and $T=\sum\lambda_iE_i$ holds on every $E_{\lambda_l}$. (a) $m_T$ divides such a polynomial, so it too has distinct roots in $\mathbb{F}$. (b) $m_{T|_W}\mid m_T$ (Theorem 9.14(1)), and a divisor of a product of distinct linear factors is again one. (c) $m_T$ divides $x^r$ and has distinct roots, so $m_T=x$ and $T=0$. ∎

🧭 Problem-solving strategy — is $A$ diagonalisable?
  1. Does $p_A$ split over the field? If not (e.g. a factor $x^2+1$ over $\mathbb{R}$), the answer is no.
  2. Distinct eigenvalues? Then yes (Theorem 9.13).
  3. Repeated eigenvalue $\lambda$: compare $g(\lambda)=n-\operatorname{rank}(A-\lambda I)$ with $a(\lambda)$; or test whether $\prod(A-\lambda_iI)=0$ over the distinct eigenvalues.
  4. Given a polynomial identity ($A^2=A$, $A^3=A$, $A^k=I$, $A^2-3A+2I=0$): factor it. Distinct roots in the field $\Rightarrow$ yes (Theorem 9.16(a)). A repeated root decides nothing until you know $m_A$.
  5. Symmetric real matrices are always diagonalisable (§10.2); nonzero nilpotent ones never are.
✏️ Example 3 — Deciding diagonalisability from a polynomial identity
Decide whether $A\in M_n(\mathbb{R})$ must be diagonalisable over $\mathbb{R}$ (or over $\mathbb{C}$) if: (a) $A^2-3A+2I=0$; (b) $A^3=A$; (c) $A^2=0$, $A\ne0$; (d) $A^4=I$; (e) $A=\begin{pmatrix}1&1&0\\0&1&0\\0&0&2\end{pmatrix}$.
(a), (b) $x^2-3x+2=(x-1)(x-2)$ and $x^3-x=x(x-1)(x+1)$ have distinct real roots, so $A$ is diagonalisable over $\mathbb{R}$ (Theorem 9.16(a)), with eigenvalues in $\{1,2\}$, resp. $\{0,1,-1\}$.
(c) $m_A\mid x^2$ and $m_A\ne x$ (as $A\ne0$), so $m_A=x^2$: not diagonalisable over any field.
(d) $x^4-1=(x-1)(x+1)(x-i)(x+i)$ has distinct roots in $\mathbb{C}$: always diagonalisable over $\mathbb{C}$. Over $\mathbb{R}$ only if $x^2+1\nmid m_A$: the rotation $R$ by $90^\circ$ satisfies $R^4=I$ and is not diagonalisable over $\mathbb{R}$, whereas $\operatorname{diag}(1,-1)$ is.
(e) $p_A=(x-1)^2(x-2)$; candidates for $m_A$ are $(x-1)(x-2)$ and $p_A$. $(A-I)(A-2I)=\begin{pmatrix}0&1&0\\0&0&0\\0&0&1\end{pmatrix}\begin{pmatrix}-1&1&0\\0&-1&0\\0&0&0\end{pmatrix}=\begin{pmatrix}0&-1&0\\0&0&0\\0&0&0\end{pmatrix}\ne0$, so $m_A=(x-1)^2(x-2)$ has a repeated factor: not diagonalisable.
✏️ Example 4 — A primary decomposition, with its projections
Let $A=\begin{pmatrix}2&1&-1\\-1&1&1\\0&1&1\end{pmatrix}$. Find $m_A$, the primary components $W_1$, $W_2$, the projections $E_1$, $E_2$ as polynomials in $A$, and the matrix of $A$ in a basis adapted to $\mathbb{R}^3=W_1\oplus W_2$.
$p_A$, $m_A$ $\operatorname{tr}A=4$; the principal $2\times2$ minors are $\{1,2\}$: $2\cdot1-1\cdot(-1)=3$, $\{1,3\}$: $2\cdot1-(-1)\cdot0=2$, $\{2,3\}$: $1\cdot1-1\cdot1=0$, so $E_2=5$; expanding along the first row, $\det A=2(1-1)-1(-1-0)+(-1)(-1-0)=2$. So $p_A=x^3-4x^2+5x-2=(x-1)^2(x-2)$. $A-I=\begin{pmatrix}1&1&-1\\-1&0&1\\0&1&0\end{pmatrix}$ has rank $2$, so $g(1)=1\lt2$ and $m_A=p_A$.
Bézout With $f_1=x-2$ (for the factor $(x-1)^2$) and $f_2=(x-1)^2$ (for $x-2$): $(x-1)^2-x(x-2)=1$. So $g_2=1$, $g_1=-x$, and $$E_2=(A-I)^2=\begin{pmatrix}0&0&0\\-1&0&1\\-1&0&1\end{pmatrix},\qquad E_1=I-E_2=-A(A-2I)=\begin{pmatrix}1&0&0\\1&1&-1\\1&0&0\end{pmatrix}.$$ Check: $E_2^2=E_2$ and $E_1E_2=0$.
Components $W_2=\operatorname{im}E_2=\ker(A-2I)=\operatorname{span}\{(0,1,1)\}$ and $W_1=\operatorname{im}E_1=\ker(A-I)^2=\operatorname{span}\{(1,0,1),(0,1,0)\}$; $\dim W_1=2=a(1)$, $\dim W_2=1=a(2)$, as Theorem 9.15 predicts.
Block form $A(1,0,1)^T=(1,0,1)^T$ and $A(0,1,0)^T=(1,1,1)^T=(1,0,1)^T+(0,1,0)^T$. In the basis $(1,0,1),(0,1,0),(0,1,1)$, $$[A]=\begin{pmatrix}1&1&0\\0&1&0\\0&0&2\end{pmatrix}:$$ on $W_1$, $A=I+N$ with $N\ne0$, $N^2=0$ (minimal polynomial $(x-1)^2$); on $W_2$, $A=2$. This is already the Jordan form of §9.3.

Triangular Forms and Schur's Theorem

📘 Definition 9.8 — Triangularisable operator; flag
$T$ is triangularisable if $[T]_{\mathcal{B}}$ is upper triangular for some ordered basis $\mathcal{B}=(b_1,\dots,b_n)$. Equivalently, there is a flag of $T$-invariant subspaces $\{0\}=V_0\subset V_1\subset\cdots\subset V_n=V$ with $\dim V_j=j$ (take $V_j=\operatorname{span}\{b_1,\dots,b_j\}$: upper triangularity says exactly $Tb_j\in V_j$).
📐 Theorem 9.17 — Triangularisation
$T\in\mathcal{L}(V)$ is triangularisable over $\mathbb{F}$ if and only if $p_T$ splits over $\mathbb{F}$ (equivalently, $m_T$ splits). The diagonal of any triangular matrix of $T$ lists the eigenvalues with their algebraic multiplicities. In particular, every $A\in M_n(\mathbb{C})$ is similar to an upper triangular matrix.

Proof. ($\Rightarrow$) If $[T]_{\mathcal{B}}$ is upper triangular with diagonal $t_{11},\dots,t_{nn}$, then $p_T=\prod(x-t_{jj})$ splits (Theorem 9.2(5)), and the $t_{jj}$ are the eigenvalues with multiplicity. ($\Leftarrow$) Induction on $n$; $n=1$ is trivial. Since $p_T$ splits it has a root $\lambda_1\in\mathbb{F}$; let $b_1$ be an eigenvector and $W=\operatorname{span}\{b_1\}$, which is invariant. By Theorem 9.14(1), $p_T=(x-\lambda_1)\,p_{\overline{T}}$, where $\overline{T}$ acts on $V/W$ of dimension $n-1$; so $p_{\overline{T}}$ also splits. By induction $V/W$ has a basis $b_2+W,\dots,b_n+W$ in which $\overline{T}$ is upper triangular: $\overline{T}(b_j+W)\in\operatorname{span}\{b_2+W,\dots,b_j+W\}$. Lifting, $Tb_j\in\operatorname{span}\{b_1,b_2,\dots,b_j\}$ for $j\ge2$, and $Tb_1=\lambda_1b_1$. The vectors $b_1,\dots,b_n$ form a basis of $V$ (their classes together with $W$ span), and $[T]$ is upper triangular. The minimal polynomial splits iff the characteristic one does, because they have the same irreducible factors (Theorem 9.8). ∎

A second proof of Cayley–Hamilton (over $\mathbb{C}$). With a flag $V_0\subset\cdots\subset V_n$ and diagonal entries $\lambda_1,\dots,\lambda_n$, $(T-\lambda_jI)V_j\subseteq V_{j-1}$, because $Tb_j-\lambda_jb_j\in V_{j-1}$. Hence $(T-\lambda_1I)(T-\lambda_2I)\cdots(T-\lambda_nI)$ maps $V=V_n$ into $V_{n-1}$, then into $V_{n-2}$, …, and finally into $V_0=\{0\}$. That product is $p_T(T)$. The same flag argument proves the multiplicity part of Theorem 9.9: $f(T)$ is triangular with diagonal $f(\lambda_j)$.

📐 Theorem 9.18 — Schur's unitary triangularisation
For every $A\in M_n(\mathbb{C})$ there is a unitary $U$ ($U^*U=I$, where $U^*=\overline{U}^{\,T}$) such that $U^*AU=T$ is upper triangular; the diagonal of $T$ consists of the eigenvalues of $A$. If $A\in M_n(\mathbb{R})$ has only real eigenvalues, $U$ can be taken real orthogonal ($Q^TQ=I$). Consequences:
  1. (Schur's inequality) $\sum_i|\lambda_i|^2\le\sum_{i,j}|a_{ij}|^2=\operatorname{tr}(A^*A)$, with equality if and only if $T$ is diagonal, i.e. $A$ is unitarily diagonalisable;
  2. the diagonalisable matrices are dense in $M_n(\mathbb{C})$: every $A$ is a limit of matrices with $n$ distinct eigenvalues.

Proof. Induction on $n$, as in Theorem 9.17 but keeping orthonormality. Let $\lambda_1$ be an eigenvalue and $u_1$ a unit eigenvector. Extend $u_1$ to an orthonormal basis $u_1,\dots,u_n$ of $\mathbb{C}^n$ (Gram–Schmidt, §10.1) and let $U_1$ be the unitary matrix with these columns. The first column of $U_1^*AU_1$ is $U_1^*Au_1=\lambda_1U_1^*u_1=\lambda_1e_1$, so

$$U_1^*AU_1=\begin{pmatrix}\lambda_1&*\\0&A_1\end{pmatrix},\qquad A_1\in M_{n-1}(\mathbb{C}).$$

By induction $U_2^*A_1U_2=T_1$ is upper triangular for a unitary $U_2$; then $U=U_1\operatorname{diag}(1,U_2)$ is unitary and $U^*AU=\begin{pmatrix}\lambda_1&*\\0&T_1\end{pmatrix}$ is upper triangular. Its diagonal lists the eigenvalues (Theorem 9.17). If $A$ is real with real eigenvalues, the eigenvector $u_1$ can be chosen real, Gram–Schmidt stays in $\mathbb{R}^n$, and $A_1$ is real with real eigenvalues, so the same induction runs with real orthogonal matrices.

(1) Since $A^*A=UT^*U^*UTU^*=UT^*TU^*$ and the trace is a similarity invariant,

$$\operatorname{tr}(A^*A)=\operatorname{tr}(T^*T)=\sum_{i\le j}|t_{ij}|^2=\sum_i|\lambda_i|^2+\sum_{i\lt j}|t_{ij}|^2,$$

which is $\ge\sum|\lambda_i|^2$, with equality iff all $t_{ij}$ ($i\lt j$) vanish. (2) Given $\varepsilon\gt0$, change the diagonal entries of $T$ by amounts $\delta_i$ with $|\delta_i|\lt\varepsilon$ so that $t_{ii}+\delta_i$ are distinct, and put $A_\varepsilon=U(T+\operatorname{diag}(\delta_i))U^*$. It has $n$ distinct eigenvalues, so it is diagonalisable (Theorem 9.13), and $\operatorname{tr}\big((A_\varepsilon-A)^*(A_\varepsilon-A)\big)=\sum|\delta_i|^2\lt n\varepsilon^2$. ∎

✏️ Example 5 — A triangular form and a Schur form
(a) The matrix $C=\begin{pmatrix}3&1&-1\\3&2&-3\\4&1&-2\end{pmatrix}$ of §9.1 (eigenvalues $2,2,-1$, with $g(2)=1$) is not diagonalisable. Find an invertible $P$ with $P^{-1}CP$ upper triangular. (b) Find a real orthogonal $Q$ with $Q^TAQ$ upper triangular for $A=\begin{pmatrix}1&2\\-2&5\end{pmatrix}$, and check Schur's inequality.
(a) The flag Start with the eigenvector $b_1=(1,0,1)$ for $2$. Next we need $b_2$ with $Cb_2\in\operatorname{span}\{b_1,b_2\}$; the natural choice is a solution of $(C-2I)b_2=b_1$: $b_2=(0,1,0)$ works, since the second column of $C-2I$ is $(1,0,1)^T$. Finally take the eigenvector $b_3=(0,1,1)$ for $-1$.
(a) The matrix $Cb_1=2b_1$, $Cb_2=b_1+2b_2$, $Cb_3=-b_3$. With $P=\begin{pmatrix}1&0&0\\0&1&1\\1&0&1\end{pmatrix}$, $$P^{-1}CP=\begin{pmatrix}2&1&0\\0&2&0\\0&0&-1\end{pmatrix}.$$ (Upper triangular — in fact the Jordan form, §9.3.)
(b) Schur $p_A=x^2-6x+9=(x-3)^2$; $(A-3I)v=0$ gives the unit eigenvector $u_1=\tfrac1{\sqrt2}(1,1)$. Complete it to an orthonormal basis with $u_2=\tfrac1{\sqrt2}(-1,1)$: $Q=\tfrac1{\sqrt2}\begin{pmatrix}1&-1\\1&1\end{pmatrix}$, and $$Q^TAQ=\begin{pmatrix}3&4\\0&3\end{pmatrix}.$$
(b) Check $\sum|a_{ij}|^2=1+4+4+25=34$ and $\sum|\lambda_i|^2+|t_{12}|^2=9+9+16=34$. ✓ The gap $16\gt0$ says that $A$ is not orthogonally diagonalisable (it is not even diagonalisable).
Interactive 2 — Schur Triangularisation, One Orthogonal Step at a Time

Each step finds a unit eigenvector of the remaining block, and a reflection $H$ (an orthogonal matrix) moves it to the first basis vector of that block; then $A\mapsto HAH$. The entries below the diagonal (amber frame) are cleared column by column. On the right, $\sum|a_{ij}|^2$ is split into diagonal, upper and lower parts: the total never changes, and at the end the diagonal part is $\sum\lambda_i^2$ — the leftover upper part is the gap in Schur's inequality, zero exactly for the symmetric preset.

step—
diagonal so far—
Σ (below diagonal)²—
Σ aᵢⱼ² (invariant)—
Schur gap Σ λ² vs total—

Simultaneous Diagonalisation and Triangularisation

📘 Definition 9.9 — Simultaneously diagonalisable (triangularisable)
A family $\mathcal{F}\subseteq\mathcal{L}(V)$ is simultaneously diagonalisable (resp. simultaneously triangularisable) if one ordered basis $\mathcal{B}$ makes $[T]_{\mathcal{B}}$ diagonal (resp. upper triangular) for every $T\in\mathcal{F}$; for matrices, one invertible $P$ makes every $P^{-1}AP$ diagonal (upper triangular).
📐 Theorem 9.19 — Commuting diagonalisable operators
Let $S,T\in\mathcal{L}(V)$ be diagonalisable. Then $S$ and $T$ are simultaneously diagonalisable if and only if $ST=TS$. More generally, any family of diagonalisable operators that commute pairwise is simultaneously diagonalisable. In particular, if $S,T$ are diagonalisable and commute, then $S+T$, $ST$ and every polynomial $f(S,T)$ are diagonalisable.

Proof. ($\Rightarrow$) Diagonal matrices commute, so $[ST]=[S][T]=[T][S]=[TS]$ in the common basis. ($\Leftarrow$) Let $\lambda_1,\dots,\lambda_k$ be the distinct eigenvalues of $T$, so $V=E_{\lambda_1}(T)\oplus\cdots\oplus E_{\lambda_k}(T)$ (Theorem 9.12). Each $E_{\lambda_i}(T)$ is $S$-invariant because $S$ commutes with $T$ (Theorem 9.14(3)), and $S|_{E_{\lambda_i}(T)}$ is diagonalisable by Theorem 9.16(b). Choose a basis of each $E_{\lambda_i}(T)$ consisting of eigenvectors of $S$; its vectors are also eigenvectors of $T$ (everything in $E_{\lambda_i}(T)$ is). The union of these bases is a basis of $V$ of common eigenvectors. For a commuting family, induct on $\dim V$: if every member is a scalar there is nothing to do; otherwise some $T$ in the family has at least two eigenvalues, every eigenspace $E_{\lambda_i}(T)$ is invariant under the whole family and has smaller dimension, and the restrictions are diagonalisable and commute, so by induction each $E_{\lambda_i}(T)$ has a basis of common eigenvectors. The last statement: in the common basis $S+T$, $ST$, $f(S,T)$ are diagonal. ∎

📐 Theorem 9.20 — Commuting operators have a common eigenvector and a common triangular form
Let $\mathcal{F}\subseteq\mathcal{L}(V)$ be a family of pairwise commuting operators on a finite-dimensional complex space (more generally, over any $\mathbb{F}$ over which every $p_T$, $T\in\mathcal{F}$, splits). Then (1) the members of $\mathcal{F}$ have a common eigenvector; (2) $\mathcal{F}$ is simultaneously triangularisable — over $\mathbb{C}$, even by one unitary matrix. Hence for commuting $A,B\in M_n(\mathbb{C})$ the eigenvalues can be listed as $\lambda_i$ for $A$ and $\mu_i$ for $B$ so that $A+B$ has eigenvalues $\lambda_i+\mu_i$ and $AB$ has eigenvalues $\lambda_i\mu_i$.

Proof (with every key step). (1) The family spans a subspace of the finite-dimensional space $\mathcal{L}(V)$, so it has a finite basis $T_1,\dots,T_r$ taken from $\mathcal{F}$, and a common eigenvector of $T_1,\dots,T_r$ is one of every member (each member is a combination of them). Let $W_0=V$. $T_1|_{W_0}$ has an eigenvalue $\lambda_1$ (its characteristic polynomial splits); $W_1=\ker(T_1-\lambda_1I)\ne\{0\}$ is invariant under every $T_j$ (they commute with $T_1$). The restriction $T_2|_{W_1}$ has characteristic polynomial dividing $p_{T_2}$ (Theorem 9.14(1)), so it splits and has an eigenvalue $\lambda_2$; $W_2=W_1\cap\ker(T_2-\lambda_2I)\ne\{0\}$ is again invariant under all $T_j$. After $r$ steps, every nonzero vector of $W_r$ is a common eigenvector. (2) Induction on $n$ as in Theorem 9.17: take a common eigenvector $b_1$, pass to $V/\operatorname{span}\{b_1\}$, on which the induced operators still commute and have split characteristic polynomials, triangularise them simultaneously by induction, and lift. For the unitary version, run the proof of Theorem 9.18 with a common unit eigenvector $u_1$; the blocks $A_1,B_1,\dots$ still commute because $U_1^*AU_1$ and $U_1^*BU_1$ commute and are block triangular. Finally, if $U^*AU$ and $U^*BU$ are upper triangular with diagonals $(\lambda_i)$, $(\mu_i)$, then $U^*(A+B)U$ and $U^*ABU$ are upper triangular with diagonals $(\lambda_i+\mu_i)$, $(\lambda_i\mu_i)$. ∎

✏️ Example 6 — Simultaneous diagonalisation, following the proof
Let $A=\begin{pmatrix}1&0&0\\-1&1&1\\-1&0&2\end{pmatrix}$ and $B=\begin{pmatrix}1&1&-1\\0&1&0\\0&1&0\end{pmatrix}$. Show that $A$ and $B$ are diagonalisable and commute, and find a single $P$ that diagonalises both. Is $B$ a polynomial in $A$?
Eigenspaces of $A$ $A-I=\begin{pmatrix}0&0&0\\-1&0&1\\-1&0&1\end{pmatrix}$ has rank $1$: $E_1(A)=\{v_1=v_3\}=\operatorname{span}\{(0,1,0),(1,0,1)\}$. Since $\operatorname{tr}A=4$, the third eigenvalue is $2$, with $E_2(A)=\operatorname{span}\{(0,1,1)\}$. So $A$ is diagonalisable, but its eigenspace $E_1(A)$ is a plane, and not every basis of that plane will diagonalise $B$.
Commuting Direct multiplication gives $AB=BA=\begin{pmatrix}1&1&-1\\-1&1&1\\-1&1&1\end{pmatrix}$. ($B$ is diagonalisable too: $B^2=B$, Theorem 9.16(a).)
Split $E_1(A)$ by $B$ $B$ maps $E_1(A)$ into itself: $B(0,1,0)^T=(1,1,1)^T=(0,1,0)^T+(1,0,1)^T$ and $B(1,0,1)^T=0$. On this plane $B$ has the eigenvectors $(1,0,1)$ (eigenvalue $0$) and $(1,1,1)$ (eigenvalue $1$). Also $B(0,1,1)^T=(0,1,1)^T$.
Common basis $P=\begin{pmatrix}1&1&0\\0&1&1\\1&1&1\end{pmatrix}$ (columns $(1,0,1),(1,1,1),(0,1,1)$) gives $P^{-1}AP=\operatorname{diag}(1,1,2)$ and $P^{-1}BP=\operatorname{diag}(0,1,1)$. $B$ is not a polynomial in $A$: any $f(A)$ acts as the single scalar $f(1)$ on $E_1(A)$, whereas $B$ takes the values $0$ and $1$ there. (When $A$ has $n$ distinct eigenvalues, every matrix commuting with $A$ is a polynomial in $A$ — Exercise IV.3.)
✏️ Example 7 — Prove or disprove
Let $A,B\in M_n(\mathbb{C})$. (a) If $AB=BA$ and $A$ is diagonalisable, then $B$ is diagonalisable. (b) If $A$ and $B$ are diagonalisable, then $A+B$ is diagonalisable. (c) If $A$ and $B$ are diagonalisable and $AB=BA$, then $A+B$ and $AB$ are diagonalisable. (d) If $A^2$ is diagonalisable, then $A$ is diagonalisable. (e) If $A$ is invertible and $A^2$ is diagonalisable, then $A$ is diagonalisable.
(a) False $A=I$ commutes with everything, including $B=\begin{pmatrix}0&1\\0&0\end{pmatrix}$.
(b) False $A=\begin{pmatrix}1&1\\0&0\end{pmatrix}$ (eigenvalues $1,0$) and $B=\begin{pmatrix}-1&0\\0&0\end{pmatrix}$ are diagonalisable, but $A+B=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ is not.
(c) True Theorem 9.19: in a common eigenbasis both $A+B$ and $AB$ are diagonal.
(d) False $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ has $A^2=0$, which is diagonal.
(e) True $m_{A^2}=\prod_{i=1}^k(x-\mu_i)$ with distinct $\mu_i$, all nonzero because $A^2$ is invertible. Then $A$ is annihilated by $f(x)=\prod_i(x^2-\mu_i)$. Choosing square roots $\nu_i$ of $\mu_i$, $f=\prod_i(x-\nu_i)(x+\nu_i)$, and its $2k$ roots $\pm\nu_i$ are distinct: $\nu_i\ne-\nu_i$ as $\mu_i\ne0$, and $\pm\nu_i\ne\pm\nu_j$ for $i\ne j$ since their squares differ. By Theorem 9.16(a), $A$ is diagonalisable.
⚠️ Counterexamples and pitfalls
  1. Real eigenvalues do not imply diagonalisable: $\begin{pmatrix}1&1\\0&1\end{pmatrix}$, or any nonzero nilpotent matrix.
  2. Diagonalisable depends on the field: the $90^\circ$ rotation is diagonalisable over $\mathbb{C}$, not over $\mathbb{R}$; $\begin{pmatrix}0&1\\1&0\end{pmatrix}$ over $\mathbb{R}$, not over $\mathbb{F}_2$.
  3. Distinct eigenvalues is sufficient, not necessary: $I_n$, or $D$ of Example 1.
  4. Invertible $\ne$ diagonalisable, in both directions: $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ is invertible but not diagonalisable; $\operatorname{diag}(1,0)$ is diagonalisable but singular.
  5. A repeated root in an annihilating polynomial decides nothing: $A^2=2A-I$ holds for $I$ (diagonalisable) and for the shear (not). Only the minimal polynomial decides.
  6. Sums and products of diagonalisable matrices need not be diagonalisable (Example 7(b); $\operatorname{diag}(1,0)\begin{pmatrix}0&1\\1&0\end{pmatrix}=\begin{pmatrix}0&1\\0&0\end{pmatrix}$) — unless they commute.
  7. $A^2$ diagonalisable does not give $A$ diagonalisable ($A$ nilpotent), but it does when $A$ is invertible (over $\mathbb{C}$).
  8. Commuting does not mean "same eigenvectors": $I$ commutes with every matrix. Commuting diagonalisable matrices share some eigenbasis, not every eigenvector.
  9. Triangular over $\mathbb{R}$ needs real eigenvalues: a real matrix with a non-real eigenvalue is not similar over $\mathbb{R}$ to any triangular matrix, but it is over $\mathbb{C}$.
  10. Triangular forms are not unique (the order of the diagonal and the entries above it change with the basis); only the diagonal multiset is determined.
  11. Non-commuting matrices need not share an eigenvector: $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ and $\begin{pmatrix}1&0\\1&1\end{pmatrix}$ have the eigenvectors $e_1$ and $e_2$ only, so they cannot be triangularised simultaneously.
  12. Schur's inequality is an equality only for normal matrices: for a real matrix with real eigenvalues, $\sum a_{ij}^2=\sum\lambda_i^2$ forces $A$ to be orthogonally diagonalisable (symmetric).

Exercises

I
Diagonalising and using diagonal forms.
  • 1 Diagonalise $\begin{pmatrix}1&2\\2&1\end{pmatrix}$.
  • 2 Diagonalise $\begin{pmatrix}2&0&0\\3&2&-3\\3&0&-1\end{pmatrix}$.
  • 3 Show that $\begin{pmatrix}1&0&1\\0&1&1\\1&-1&2\end{pmatrix}$ is not diagonalisable over any field containing its eigenvalues.
  • 4 For which real $a,b$ is $\begin{pmatrix}1&a&b\\0&1&0\\0&0&2\end{pmatrix}$ diagonalisable?
  • 5 For $A=\begin{pmatrix}4&-2\\1&1\end{pmatrix}$ find a formula for $A^n$.
  • 6 Find a real matrix $B$ with $B^2=\begin{pmatrix}5&4\\4&5\end{pmatrix}$.
  • 7 Applied. Each year $80\%$ of the customers of company X stay and $20\%$ switch to Y, while $30\%$ of Y's customers switch to X. With $M=\begin{pmatrix}0.8&0.3\\0.2&0.7\end{pmatrix}$ and $x_{n+1}=Mx_n$, find $M^n$ and the long-run market shares.
  • 8 Solve $x'=Ax$, $x(0)=(1,0)^T$, for $A=\begin{pmatrix}1&2\\2&1\end{pmatrix}$ by diagonalising $A$.
1 $p=x^2-2x-3=(x-3)(x+1)$; eigenvectors $(1,1)$, $(1,-1)$. $P=\begin{pmatrix}1&1\\1&-1\end{pmatrix}$, $P^{-1}AP=\operatorname{diag}(3,-1)$.
2 $p=(x-2)^2(x+1)$. $A-2I=\begin{pmatrix}0&0&0\\3&0&-3\\3&0&-3\end{pmatrix}$ has rank $1$: $E_2=\operatorname{span}\{(0,1,0),(1,0,1)\}$. $E_{-1}=\operatorname{span}\{(0,1,1)\}$. $P=\begin{pmatrix}0&1&0\\1&0&1\\0&1&1\end{pmatrix}$, $P^{-1}AP=\operatorname{diag}(2,2,-1)$.
3 $\operatorname{tr}=4$, $E_2=1+1+3=5$, $\det=2$: $p=x^3-4x^2+5x-2=(x-1)^2(x-2)$. $A-I=\begin{pmatrix}0&0&1\\0&0&1\\1&-1&1\end{pmatrix}$ has rank $2$, so $g(1)=1\lt a(1)=2$: not diagonalisable, whatever the field (the rank computation is field-independent here).
4 Eigenvalues $1$ ($a=2$) and $2$. $A-I=\begin{pmatrix}0&a&b\\0&0&0\\0&0&1\end{pmatrix}$ has rank $2$ if $a\ne0$ and rank $1$ if $a=0$. So diagonalisable iff $a=0$ ($b$ arbitrary).
5 Eigenvalues $2,3$ (§9.1, Exercise I.1), so $m_A=(x-2)(x-3)$ and $A^n=\alpha A+\beta I$ with $2\alpha+\beta=2^n$, $3\alpha+\beta=3^n$: $$A^n=(3^n-2^n)A+(3\cdot2^n-2\cdot3^n)I.$$ Check $n=1$: $A+0\cdot I$. ✓
6 Eigenvalues $9$ (vector $(1,1)$) and $1$ (vector $(1,-1)$). Take $B=P\operatorname{diag}(3,1)P^{-1}=\begin{pmatrix}2&1\\1&2\end{pmatrix}$; indeed $B^2=\begin{pmatrix}5&4\\4&5\end{pmatrix}$. (Other sign choices give three more square roots.)
7 Eigenvalues $1$ (vector $(3,2)$) and $0.5$ (vector $(1,-1)$; $\operatorname{tr}M=1.5$). With $\Pi=\begin{pmatrix}0.6&0.6\\0.4&0.4\end{pmatrix}$ (the projection onto $E_1$ along $E_{0.5}$, Theorem 9.16), $M^n=\Pi+0.5^n(I-\Pi)\to\Pi$. Long run: $60\%$ for X, $40\%$ for Y, from any starting split.
8 $(1,0)=\tfrac12(1,1)+\tfrac12(1,-1)$, and each eigen-component evolves by $e^{\lambda t}$: $$x(t)=\tfrac12e^{3t}\begin{pmatrix}1\\1\end{pmatrix}+\tfrac12e^{-t}\begin{pmatrix}1\\-1\end{pmatrix}.$$
II
The minimal polynomial and diagonalisability.
  • 1 Show that every $A$ with $A^2-3A+2I=0$ is diagonalisable, and list the possible values of $\operatorname{tr}A$ for $A\in M_3(\mathbb{R})$.
  • 2 Prove that a matrix that is both nilpotent and diagonalisable is $0$.
  • 3 Prove that $A\in M_n(\mathbb{C})$ with $A^k=I$ is diagonalisable. Show that over $\mathbb{F}_2$ the matrix $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ satisfies $A^2=I$ but is not diagonalisable. Which hypothesis fails?
  • 4 Let $A\in M_n(\mathbb{R})$ with $A^3=A$. Prove that $A$ is diagonalisable over $\mathbb{R}$ and that $\operatorname{rank}A=\operatorname{tr}(A^2)$.
  • 5 Let $T$ be diagonalisable and $W$ a $T$-invariant subspace. Prove that $W=\bigoplus_\lambda(W\cap E_\lambda)$.
  • 6 Let $q$ be a prime power. Prove that $A\in M_n(\mathbb{F}_q)$ is diagonalisable over $\mathbb{F}_q$ if and only if $A^q=A$.
  • 7 Let $u,v\in\mathbb{R}^n$ be nonzero. Prove that $uv^T$ is diagonalisable if and only if $v^Tu\ne0$.
  • 8 Show that $A\in M_n(\mathbb{C})$ is diagonalisable iff $\operatorname{rank}(A-\lambda I)=\operatorname{rank}(A-\lambda I)^2$ for every eigenvalue $\lambda$.
1 $x^2-3x+2=(x-1)(x-2)$ has distinct roots, so $A$ is diagonalisable with eigenvalues in $\{1,2\}$. In $M_3$: $\operatorname{tr}A\in\{3,4,5,6\}$ (from $1+1+1$ up to $2+2+2$).
2 Theorem 9.16(c): $m_A$ divides $x^r$ and has distinct roots, so $m_A=x$ and $A=0$. (Or: $A=PDP^{-1}$ with $D$ nilpotent diagonal, hence $D=0$.)
3 $x^k-1$ has $k$ distinct roots in $\mathbb{C}$ (its derivative $kx^{k-1}$ has no common root with it), so Theorem 9.16(a) applies. Over $\mathbb{F}_2$, $x^2-1=(x-1)^2$ has a repeated root because the derivative $2x$ is identically $0$: the hypothesis "distinct roots" fails when the characteristic of the field divides $k$. Indeed $\begin{pmatrix}1&1\\0&1\end{pmatrix}^2=\begin{pmatrix}1&2\\0&1\end{pmatrix}=I$ in $\mathbb{F}_2$, and $g(1)=1$.
4 $x^3-x=x(x-1)(x+1)$, distinct real roots: diagonalisable, $A=P\operatorname{diag}(\varepsilon_i)P^{-1}$ with $\varepsilon_i\in\{0,\pm1\}$. Then $\operatorname{rank}A=\#\{i:\varepsilon_i\ne0\}=\sum\varepsilon_i^2=\operatorname{tr}(A^2)$.
5 $T|_W$ is diagonalisable (Theorem 9.16(b)), so $W$ is the direct sum of the eigenspaces of $T|_W$, which are exactly $W\cap E_\lambda$.
6 In $\mathbb{F}_q$ every $a$ satisfies $a^q=a$, and $x^q-x=\prod_{a\in\mathbb{F}_q}(x-a)$ has $q$ distinct roots. If $A^q=A$, Theorem 9.16(a) gives diagonalisability. Conversely $A=PDP^{-1}$ with $D$ diagonal over $\mathbb{F}_q$ gives $A^q=PD^qP^{-1}=PDP^{-1}=A$.
7 $A=uv^T$ satisfies $A^2=u(v^Tu)v^T=cA$ with $c=v^Tu$. If $c\ne0$, $x(x-c)$ has distinct roots: diagonalisable. If $c=0$, $A^2=0$ and $A\ne0$ (as $u,v\ne0$): not diagonalisable (II.2).
8 If $A$ is diagonalisable, $A-\lambda I$ is similar to a diagonal matrix $D_\lambda$, and $D_\lambda$ and $D_\lambda^2$ have the same zero pattern. Conversely, if the ranks agree then $\ker(A-\lambda I)=\ker(A-\lambda I)^2$, and then all higher kernels agree (if $(A-\lambda I)^{j+1}v=0$ then $(A-\lambda I)^{j-1}v\in\ker(A-\lambda I)^2=\ker(A-\lambda I)$, and induct — the argument of Theorem 8.29(1) in §8.2), so the generalised eigenspace $\ker(A-\lambda I)^{a(\lambda)}$ equals $E_\lambda$; by Theorem 9.15, $\dim E_\lambda=a(\lambda)$ for every $\lambda$, and Theorem 9.12(5) applies.
III
Invariant subspaces, primary decomposition, triangular forms.
  • 1 Find all invariant subspaces of $\mathbb{R}^2$ for (a) $\operatorname{diag}(2,3)$; (b) the rotation by $90^\circ$; (c) $\begin{pmatrix}0&1\\0&0\end{pmatrix}$.
  • 2 Prove that if $W$ is $T$-invariant and $T$ is invertible (finite dimension), then $W$ is $T^{-1}$-invariant. Give an infinite-dimensional counterexample.
  • 3 For $A=\begin{pmatrix}1&0&1\\0&1&1\\1&-1&2\end{pmatrix}$ find the primary components $W_1=\ker(A-I)^2$, $W_2=\ker(A-2I)$ and the projection onto $W_2$ as a polynomial in $A$.
  • 4 Find an invertible $P$ with $P^{-1}AP$ upper triangular for $A=\begin{pmatrix}1&1\\-1&3\end{pmatrix}$.
  • 5 Find a real orthogonal $Q$ with $Q^TAQ$ upper triangular for $A=\begin{pmatrix}3&1\\-1&1\end{pmatrix}$.
  • 6 Prove that a real matrix with a non-real eigenvalue is not similar over $\mathbb{R}$ to an upper triangular matrix.
  • 7 Use Schur's theorem to prove that $A\in M_n(\mathbb{C})$ is nilpotent iff all its eigenvalues are $0$, and then that $\operatorname{tr}(A^k)=0$ for $k=1,\dots,n$ implies $A$ nilpotent.
  • 8 Show that the diagonalisable matrices are not dense in $M_2(\mathbb{R})$: find an open set of real $2\times2$ matrices none of which is diagonalisable over $\mathbb{R}$.
1 A 1-dimensional invariant subspace is spanned by an eigenvector. (a) $\{0\}$, the two axes, $\mathbb{R}^2$. (b) Only $\{0\}$ and $\mathbb{R}^2$ (no real eigenvector). (c) $\{0\}$, $\operatorname{span}\{e_1\}$, $\mathbb{R}^2$.
2 $T|_W$ is injective, hence (finite dimension) bijective onto $W$: for $w\in W$, $w=Tw'$ with $w'\in W$, so $T^{-1}w=w'\in W$. Counterexample: on two-sided sequences $(x_k)_{k\in\mathbb{Z}}$ let $T$ be the shift $(Tx)_k=x_{k-1}$, invertible, and $W=\{x:x_k=0\text{ for }k\lt0\}$; $T(W)\subseteq W$ but $T^{-1}(W)\not\subseteq W$.
3 $p=(x-1)^2(x-2)=m$ (Exercise I.3). The same Bézout identity $(x-1)^2-x(x-2)=1$ gives $E_2=(A-I)^2=\begin{pmatrix}1&-1&1\\1&-1&1\\1&-1&1\end{pmatrix}$, of rank $1$. $W_2=\operatorname{span}\{(1,1,1)\}$, $W_1=\ker(A-I)^2=\{v_1-v_2+v_3=0\}=\operatorname{span}\{(1,1,0),(-1,0,1)\}$.
4 $p=(x-2)^2$, eigenvector $(1,1)$. Complete it with $e_2$: $P=\begin{pmatrix}1&0\\1&1\end{pmatrix}$, $P^{-1}AP=\begin{pmatrix}2&1\\0&2\end{pmatrix}$.
5 $p=(x-2)^2$, unit eigenvector $\tfrac1{\sqrt2}(1,-1)$. $Q=\tfrac1{\sqrt2}\begin{pmatrix}1&1\\-1&1\end{pmatrix}$ gives $Q^TAQ=\begin{pmatrix}2&2\\0&2\end{pmatrix}$. Check: $\sum a_{ij}^2=12=4+4+4$. ✓
6 If $P^{-1}AP=T$ is real upper triangular, $p_A=\prod(x-t_{ii})$ has only real roots (Theorem 9.17).
7 With $A=UTU^*$: $A^k=UT^kU^*$ and $T^k$ is triangular with diagonal $\lambda_i^k$. If all $\lambda_i=0$, $T$ is strictly upper triangular and $T^n=0$. Conversely $A^k=0$ forces $\lambda_i^k=0$. If $\operatorname{tr}A^k=\sum\lambda_i^k=0$ for $k\le n$, Newton's identities give $e_1=\cdots=e_n=0$, so $p_A=x^n$ and all $\lambda_i=0$.
8 $\Delta(A)=(\operatorname{tr}A)^2-4\det A$ is a continuous function of the entries, and $\Delta\lt0$ means non-real eigenvalues. So $\{A:\Delta(A)\lt0\}$ is open, nonempty (it contains the rotation, $\Delta=-4$), and contains no matrix diagonalisable over $\mathbb{R}$.
IV
Commuting matrices and simultaneous forms.
  • 1 Show that $A=\begin{pmatrix}1&1\\1&1\end{pmatrix}$ and $B=\begin{pmatrix}2&-1\\-1&2\end{pmatrix}$ commute, and diagonalise them simultaneously.
  • 2 Show that $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ and $\begin{pmatrix}1&0\\1&1\end{pmatrix}$ are each triangularisable, but not simultaneously.
  • 3 Let $A\in M_n(\mathbb{F})$ have $n$ distinct eigenvalues in $\mathbb{F}$ and let $AB=BA$. Prove that $B$ is diagonalisable in every eigenbasis of $A$ and that $B$ is a polynomial in $A$.
  • 4 $A,B\in M_2(\mathbb{C})$ commute, $A$ has eigenvalues $1,2$ and $B$ has eigenvalues $3,4$. What can the eigenvalues of $A+B$ be? What if $A$ and $B$ need not commute?
  • 5 Using Theorem 9.20, prove that the sum and the product of two commuting nilpotent matrices are nilpotent.
  • 6 Find two diagonalisable $2\times2$ matrices whose product is not diagonalisable.
1 $AB=\begin{pmatrix}1&1\\1&1\end{pmatrix}=BA$. Common eigenvectors $(1,1)$ ($A$: $2$, $B$: $1$) and $(1,-1)$ ($A$: $0$, $B$: $3$). $P=\begin{pmatrix}1&1\\1&-1\end{pmatrix}$: $P^{-1}AP=\operatorname{diag}(2,0)$, $P^{-1}BP=\operatorname{diag}(1,3)$.
2 Each is already triangular (upper and lower; the second becomes upper triangular in the basis $(e_2,e_1)$). A common triangular form would need a common eigenvector (the first basis vector), but the only eigen-directions are $e_1$ and $e_2$ respectively. (They do not commute.)
3 Each eigenspace $E_{\lambda_i}(A)$ is $1$-dimensional and $B$-invariant (Theorem 9.14(3)), so every eigenvector of $A$ is an eigenvector of $B$, say $Bv_i=\mu_iv_i$. By Lagrange interpolation choose $f$ with $f(\lambda_i)=\mu_i$; then $f(A)v_i=\mu_iv_i=Bv_i$ for all $i$, so $B=f(A)$.
4 Commuting: with a common triangular form the eigenvalues pair up, giving $\{1+3,2+4\}=\{4,6\}$ or $\{1+4,2+3\}=\{5,5\}$. Without commuting only $\operatorname{tr}(A+B)=10$ survives: $A=\operatorname{diag}(1,2)$ and $B=\begin{pmatrix}5&1\\-2&2\end{pmatrix}$ (trace $7$, determinant $12$, eigenvalues $3,4$) give $A+B=\begin{pmatrix}6&1\\-2&4\end{pmatrix}$, with eigenvalues $5\pm i$.
5 In a common triangular form both are strictly upper triangular (all eigenvalues $0$, Exercise III.7); sums and products of strictly upper triangular matrices are strictly upper triangular, hence nilpotent.
6 $A=\operatorname{diag}(1,0)$ and $B=\begin{pmatrix}0&1\\1&0\end{pmatrix}$ (eigenvalues $\pm1$) are diagonalisable, but $AB=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ is not.
🎯 Exam corner CSIR Part BCSIR Part CGATE MA

E1. CSIR Part B GATE MA Which of the following real matrices is diagonalisable over $\mathbb{R}$? (A) $\begin{pmatrix}1&1&0\\0&1&0\\0&0&2\end{pmatrix}$; (B) $\begin{pmatrix}0&0&1\\1&0&0\\0&1&0\end{pmatrix}$; (C) $\begin{pmatrix}1&1&1\\1&1&1\\1&1&1\end{pmatrix}$; (D) $\begin{pmatrix}2&1\\-1&0\end{pmatrix}$.

Eliminate (A): $m=(x-1)^2(x-2)$ (Example 3(e)). (B): a cyclic permutation matrix with $p=x^3-1$, which has the non-real roots $e^{\pm2\pi i/3}$. (D): $p=x^2-2x+1=(x-1)^2$ and the matrix is not $I$, so $m=(x-1)^2$.
Why (C) $J_3^2=3J_3$, so $m=x(x-3)$ with distinct real roots (also: $J_3$ is symmetric).
Answer (C).

E2. CSIR Part B Let $A\in M_5(\mathbb{C})$ have characteristic polynomial $(x-1)^3(x-2)^2$ and minimal polynomial $(x-1)(x-2)^2$. Then $\dim\ker(A-I)$ is (A) $1$; (B) $2$; (C) $3$; (D) not determined by the data.

Idea Primary decomposition (Theorem 9.15).
Work The factor $x-1$ appears in $m_A$ with exponent $e=1$, so the primary component for $1$ is $W_1=\ker(A-I)^1$, and $\dim W_1=a(1)=3$. (The component for $2$, $\ker(A-2I)^2$, has dimension $2$; $\dim\ker(A-2I)$ is $1$ because $A-2I\ne0$ on it.)
Answer (C).

E3. CSIR Part B Let $A\in M_3(\mathbb{C})$ have eigenvalues $1,1,2$ and $\sum_{i,j}|a_{ij}|^2=6$. Then (A) $A$ is nilpotent; (B) $A=UDU^*$ for some unitary $U$ and diagonal $D$; (C) $A$ is not diagonalisable; (D) $A$ must be upper triangular.

Idea Schur's inequality (Theorem 9.18(1)).
Work $\sum|\lambda_i|^2=1+1+4=6=\sum|a_{ij}|^2$: equality, so the Schur form $T=U^*AU$ is diagonal and $A=UTU^*$. That rules out (C); (A) is false since the eigenvalues are nonzero; (D) fails for $A=U\operatorname{diag}(1,1,2)U^*$ with a non-diagonal unitary $U$.
Answer (B).

E4. CSIR Part C Let $A\in M_4(\mathbb{R})$ satisfy $A^4=I$. Which of the following are necessarily true? (A) $A$ is diagonalisable over $\mathbb{C}$; (B) $A$ is diagonalisable over $\mathbb{R}$; (C) every eigenvalue of $A$ (in $\mathbb{C}$) lies in $\{1,-1,i,-i\}$; (D) $\det A=\pm1$.

(A) True $x^4-1$ has four distinct complex roots (Theorem 9.16(a)).
(B) False $A=\operatorname{diag}(R,R)$ with $R=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ has $A^4=I$ but eigenvalues $\pm i$.
(C) True $m_A\mid x^4-1$, and the eigenvalues are roots of $m_A$.
(D) True $(\det A)^4=1$ and $\det A$ is real, so $\det A=\pm1$.
Answer (A), (C) and (D).

E5. CSIR Part C Let $A,B\in M_n(\mathbb{C})$ be diagonalisable with $AB=BA$. Which of the following are true? (A) $A+B$ is diagonalisable; (B) $AB$ is diagonalisable; (C) every eigenvector of $A$ is an eigenvector of $B$; (D) there is an invertible $P$ with $P^{-1}AP$ and $P^{-1}BP$ both diagonal.

(D) True Theorem 9.19.
(A), (B) True In the common eigenbasis of (D), $A+B$ and $AB$ are diagonal.
(C) False $A=I$, $B=\operatorname{diag}(1,2)$: $(1,1)$ is an eigenvector of $A$ but not of $B$.
Answer (A), (B) and (D).

E6. CSIR Part C Let $A\in M_3(\mathbb{R})$. Which of the following are true? (A) $A$ is similar over $\mathbb{R}$ to an upper triangular matrix; (B) $A$ is similar over $\mathbb{C}$ to an upper triangular matrix; (C) if all eigenvalues of $A$ are real, then $A=QTQ^T$ with $Q$ real orthogonal and $T$ upper triangular; (D) if $A$ is similar over $\mathbb{R}$ to an upper triangular matrix, then $A$ is diagonalisable over $\mathbb{C}$.

(A) False $\operatorname{diag}(R,1)$ with $R$ the $90^\circ$ rotation has eigenvalues $\pm i$ (Exercise III.6).
(B) True Theorem 9.17: $p_A$ splits over $\mathbb{C}$.
(C) True The real form of Schur's theorem (Theorem 9.18).
(D) False The nilpotent shift $\begin{pmatrix}0&1&0\\0&0&1\\0&0&0\end{pmatrix}$ is already upper triangular and is not diagonalisable over any field.
Answer (B) and (C).
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min (a third for the exercises).

  1. Diagonalisation and the minimal polynomial. Theorem 9.12 in full (it is a reorganisation of §9.1), Theorem 9.13, Example 1 on the board including the "$D^k=\alpha D+\beta I$" shortcut. Example 2 as a three-minute discussion of fields. Project Interactive 1: let $s$ run through $0$ and pause there — the eigenvectors merge, $|\det P|\to0$, and the eigenvalues collide and leave the real axis. Then invariant subspaces (Theorem 9.14) and the primary decomposition with its Bézout proof (Theorem 9.15) — allow 20 minutes; Theorem 9.16 follows in five. Example 3 as a quick-fire round.
  2. Triangular and simultaneous forms. Example 4 (projections from Bézout) on the board. Theorem 9.17 by induction on the quotient, the flag proof of Cayley–Hamilton, Schur's theorem with its inequality; project Interactive 2 for the preset $C$ (gap $\gt0$) and the symmetric preset (gap $0$). Theorem 9.19 in full, Theorem 9.20 as a sketch; Example 6 shows the proof in action. Finish with Example 7 and E5.

Discussion questions. (1) Why is "distinct eigenvalues" enough but not necessary, while "minimal polynomial with distinct roots" is both? (2) A matrix close to a non-diagonalisable one is (over $\mathbb{C}$) usually diagonalisable — so why do numerical analysts avoid diagonalising nearly defective matrices? (Interactive 1: what happens to $|\det P|$?) (3) Which statements of this section change over $\mathbb{F}_p$, and why?

Misconceptions to watch for. Concluding "diagonalisable" from an annihilating polynomial with a repeated root; forgetting to check that $p_A$ splits over the given field; believing that commuting matrices have the same eigenvectors; assuming sums or products of diagonalisable matrices are diagonalisable; confusing "triangularisable" (needs only split $p_A$) with "diagonalisable"; using $P$ with eigenvectors as rows.

Suggested homework. I(2, 4, 5, 7), II(1, 3, 4, 6, 7), III(1, 3, 5, 7), IV(1, 3, 4, 6); Exam corner E2, E4, E6.

🔬 Research–Teaching NexusResearch-tutored · optional, about 45 minutes

Research question. After an intravenous dose, why does the drug concentration in a patient's blood fall fast at first and then slowly — and what do the two speeds have to do with eigenvalues?

Why it matters. Two-compartment models are the standard tool of clinical pharmacokinetics: they set dosing intervals for antibiotics and anaesthetics in hospitals worldwide. Blood (compartment 1) exchanges the drug with tissue (compartment 2) and eliminates it through the kidneys, giving a linear system $x'=Kx$. Diagonalising $K$ splits the concentration curve into two exponentials — the fast "distribution" phase and the slow "elimination" phase.

What to do.

  1. With transfer rates $k_{12}=0.3\,\text{h}^{-1}$ (blood → tissue), $k_{21}=0.1\,\text{h}^{-1}$ (tissue → blood) and elimination $k_{10}=0.2\,\text{h}^{-1}$, write $K=\begin{pmatrix}-(k_{10}+k_{12})&k_{21}\\k_{12}&-k_{21}\end{pmatrix}$ and find its eigenvalues. Why are they real and negative? (Look at the discriminant.)
  2. Diagonalise $K$ and write the blood amount as $x_1(t)=c_1e^{\lambda_1t}+c_2e^{\lambda_2t}$ for an initial dose $x(0)=(100,0)\ \text{mg}$.
  3. Compute the two half-lives $\ln2/|\lambda_i|$ and plot $\log x_1(t)$ for $0\le t\le48$ h: identify the two straight-line phases.
  4. Explain, using Interactive 1, what would go wrong with your formula if the two eigenvalues were equal.

Tools: Python (numpy.linalg.eig, scipy.linalg.expm to check $e^{tK}$) or a spreadsheet.

Hand in: $\lambda_1$, $\lambda_2$, the formula for $x_1(t)$, the two half-lives, the semi-log plot, and a two-sentence explanation for a clinician.

Going further. Read how "method of residuals" (curve stripping) recovers the two exponentials from measured blood samples, and how the repeated-eigenvalue case produces a $te^{\lambda t}$ term — the Jordan form of §9.3 at work.

Instructor note: $K=\begin{pmatrix}-0.5&0.1\\0.3&-0.1\end{pmatrix}$, $\operatorname{tr}K=-0.6$, $\det K=0.02$, discriminant $0.36-0.08=0.28\gt0$; $\lambda_{1,2}=\tfrac{-0.6\pm\sqrt{0.28}}2\approx-0.0354,\ -0.5646\ \text{h}^{-1}$, half-lives $\approx19.6$ h and $1.23$ h. In general the discriminant is $(k_{10}+k_{12}-k_{21})^2+4k_{12}k_{21}\gt0$, so the eigenvalues are always real and distinct, and both are negative because $\operatorname{tr}K\lt0\lt\det K=k_{10}k_{21}$. For $x(0)=(100,0)$: $x_1(t)\approx12.2e^{-0.0354t}+87.8e^{-0.5646t}$ mg (the coefficients are $100\cdot\frac{\lambda_i+0.1}{\lambda_i-\lambda_j}$, from $x_1(0)=100$ and $x_1'(0)=-50$); e.g. $x_1(5)\approx15.4$ mg and $x_1(24)\approx5.2$ mg.

🧪 Self-check quiz

0/5 correct

Q1. An $n\times n$ complex matrix with $n$ distinct eigenvalues is

Q2. If $A\ne0$ and $A^2=0$, then $A$ is

Q3. A real matrix has minimal polynomial $(x-1)(x^2+1)$. It is

Q4. Schur's theorem says every $A\in M_n(\mathbb{C})$ is unitarily similar to

Q5. $A\in M_3(\mathbb{R})$ satisfies $A^2=A$ and $\operatorname{tr}A=2$. Then $\operatorname{rank}A$ is