3.2 Eigenvalues, Diagonalisation, Minimal Polynomial and Jordan Form
Eigenvalues and the characteristic polynomial, Cayley–Hamilton and the minimal polynomial, the tests that decide diagonalisability, and the Jordan canonical form read off from ranks — with the similarity-class counts that GATE MA turns into NAT questions.
Section 3.1 ended with a shift operator whose matrix, after a clever change of basis, became almost diagonal. This section asks the general question: given a linear operator on an $n$-dimensional space, in which basis is its matrix simplest? The first candidates are the directions the operator merely stretches — the eigenvectors — and the stretching factors, the eigenvalues, are the roots of a polynomial of degree $n$, the characteristic polynomial. If there are enough eigenvectors to form a basis, the matrix becomes diagonal and every computation with it (powers, inverses, polynomials, the exponential $e^{At}$ of Chapter 6) becomes a computation with $n$ numbers. If there are not, the Jordan canonical form is the next best thing: a block-diagonal matrix of eigenvalues with $1$'s just above the diagonal, determined completely by the ranks of the powers $(A-\lambda I)^k$.
GATE MA examines this material more heavily than any other part of linear algebra. Typical questions ask for an eigenvalue, a determinant or a trace of a matrix built from a given one; for the minimal polynomial of a block matrix; for which values of a parameter a matrix is diagonalisable; for the Jordan form from rank information; and — as NAT questions — for the number of similarity classes of matrices with a given characteristic or minimal polynomial. The MSQs are full of plausible statements that fail for a $2\times2$ Jordan block. Throughout, $A\in M_n(\mathbb{F})$ with $\mathbb{F}=\mathbb{R}$ or $\mathbb{C}$; everything said for matrices applies to an operator $T\in\mathcal{L}(V)$ through its matrix in any basis, by the change-of-basis formula of §3.1.
Eigenvalues, Eigenvectors and the Characteristic Polynomial
Proof. The first two statements say the same thing, and the others are equivalent to the second by the invertible matrix theorem (§3.1). ∎
- $p_A$ is monic of degree $n$, and $$p_A(x)=x^n-E_1x^{n-1}+E_2x^{n-2}-\cdots+(-1)^nE_n,$$ where $E_k$ is the sum of the $\binom nk$ principal $k\times k$ minors of $A$. In particular $E_1=\operatorname{tr}A$ and $E_n=\det A$. For $n=2$: $p_A=x^2-(\operatorname{tr}A)x+\det A$.
- Over $\mathbb{C}$, $p_A(x)=(x-\lambda_1)\cdots(x-\lambda_n)$ (eigenvalues listed with algebraic multiplicity), so $\lambda_1+\cdots+\lambda_n=\operatorname{tr}A$ and $\lambda_1\cdots\lambda_n=\det A$.
- Similar matrices have the same characteristic polynomial; so do $A$ and $A^{\mathsf T}$.
- The eigenvalues of a triangular matrix are its diagonal entries; for a block triangular matrix $\begin{pmatrix}B&C\\0&D\end{pmatrix}$, $p_A=p_Bp_D$.
- For real $A$, nonreal eigenvalues come in conjugate pairs $\lambda,\bar\lambda$ (with eigenvectors $v,\bar v$), and if $n$ is odd $A$ has a real eigenvalue.
Proof. (1) The $j$-th column of $xI-A$ is $xe_j-a_j$, where $a_j$ is the $j$-th column of $A$. The determinant is linear in each column, so it expands into $2^n$ determinants: for each set $S$ of columns choose $-a_j$ for $j\in S$ and $xe_j$ for $j\notin S$. Expanding along the columns $xe_j$ leaves $x^{n-|S|}$ times $(-1)^{|S|}$ times the principal minor of $A$ on the rows and columns in $S$. Collecting the sets with $|S|=k$ gives the coefficient $(-1)^kE_k$ of $x^{n-k}$. (2) Factor $p_A$ over $\mathbb{C}$ (fundamental theorem of algebra) and compare the coefficients of $x^{n-1}$ and $x^0$ with (1). (3) $xI-P^{-1}AP=P^{-1}(xI-A)P$, and $\det(xI-A^{\mathsf T})=\det\left((xI-A)^{\mathsf T}\right)$. (4) $xI-A$ is again (block) triangular and its determinant is the product of the diagonal entries (blocks). (5) $\overline{p_A(\lambda)}=p_A(\bar\lambda)$ because $p_A$ has real coefficients, and $\overline{Av}=A\bar v$; a real polynomial of odd degree has a real root by the intermediate value theorem. ∎
- Eigenvectors $v_1,\dots,v_k$ for distinct eigenvalues $\lambda_1,\dots,\lambda_k$ are linearly independent. More generally the sum $E_{\lambda_1}+\cdots+E_{\lambda_k}$ of the eigenspaces of distinct eigenvalues is direct, so the union of bases of the eigenspaces is independent.
- For every eigenvalue, $1\le g(\lambda)\le a(\lambda)$. Hence $\sum_\lambda g(\lambda)\le\sum_\lambda a(\lambda)\le n$, with $\sum_\lambda a(\lambda)=n$ iff $p_A$ splits into linear factors over $\mathbb{F}$ (always true over $\mathbb{C}$).
Proof. (1) Induction on $k$; $k=1$ holds because $v_1\ne0$. Suppose $\sum_{i=1}^kc_iv_i=0$. Applying $A-\lambda_kI$ kills the last term and gives $\sum_{i=1}^{k-1}c_i(\lambda_i-\lambda_k)v_i=0$; by induction $c_i(\lambda_i-\lambda_k)=0$, so $c_i=0$ for $i\lt k$ because the eigenvalues are distinct, and then $c_kv_k=0$ gives $c_k=0$. For the direct sum, if $w_1+\cdots+w_k=0$ with $w_i\in E_{\lambda_i}$, the nonzero $w_i$ would be eigenvectors for distinct eigenvalues with a vanishing sum — impossible; so all $w_i=0$. (2) $g\ge1$ since an eigenvalue has an eigenvector. Let $g=g(\lambda)$, take a basis $v_1,\dots,v_g$ of $E_\lambda$ and extend it to a basis of $\mathbb{F}^n$; with $P$ the matrix of this basis,
$$P^{-1}AP=\begin{pmatrix}\lambda I_g&C\\0&D\end{pmatrix},$$so by Theorem 3.16(3),(4) $p_A(x)=(x-\lambda)^g\,p_D(x)$, and $(x-\lambda)^g$ divides $p_A$: $g\le a(\lambda)$. The rest follows because $\deg p_A=n$. ∎
- Triangularisation. Every $A\in M_n(\mathbb{C})$ is similar to an upper triangular matrix, whose diagonal entries are the eigenvalues $\lambda_1,\dots,\lambda_n$.
- Spectral mapping. For every polynomial $f$, the eigenvalues of $f(A)$, with algebraic multiplicity, are $f(\lambda_1),\dots,f(\lambda_n)$. In particular $\operatorname{tr}A^k=\sum\lambda_i^k$, $\det f(A)=\prod f(\lambda_i)$, and if $A$ is invertible the eigenvalues of $A^{-1}$ are $1/\lambda_i$. (If $Av=\lambda v$ then $f(A)v=f(\lambda)v$ for every field.)
- $AB$ and $BA$. For $A\in M_{m\times n}$, $B\in M_{n\times m}$: $x^n\,p_{AB}(x)=x^m\,p_{BA}(x)$. So $AB$ and $BA$ have the same nonzero eigenvalues with the same algebraic multiplicities, and for square $A,B$, $p_{AB}=p_{BA}$.
- Rank-one matrices. For $u,v\in\mathbb{F}^n$: $p_{uv^{\mathsf T}}(x)=x^{n-1}\left(x-v^{\mathsf T}u\right)$. For the all-ones matrix $J\in M_n$, the eigenvalues of $aI+bJ$ are $a+nb$ (once) and $a$ ($n-1$ times).
Proof. (1) Induction on $n$. $p_A$ has a root $\lambda_1\in\mathbb{C}$, with an eigenvector $v_1$; extend to a basis and conjugate as in the proof of Theorem 3.17 to get $\begin{pmatrix}\lambda_1&*\\0&A_1\end{pmatrix}$ with $A_1\in M_{n-1}(\mathbb{C})$. If $Q^{-1}A_1Q$ is upper triangular, conjugating further by $\operatorname{diag}(1,Q)$ makes the whole matrix upper triangular. Its diagonal entries are the roots of $p_A$ by Theorem 3.16(3),(4). (2) If $P^{-1}AP=U$ is upper triangular with diagonal $\lambda_i$, then $P^{-1}f(A)P=f(U)$ is upper triangular with diagonal $f(\lambda_i)$, since sums and products of upper triangular matrices multiply their diagonals entrywise. For $A^{-1}$: $Av=\lambda v$ gives $v=\lambda A^{-1}v$, so the eigenvalues of $A^{-1}$ are the $1/\lambda$; the multiplicities agree because $p_{A^{-1}}(x)=\det(xI-A^{-1})=\det(xA-I)\det A^{-1}$ is a constant times $x^n p_A(1/x)$. (3) A direct computation shows
$$\begin{pmatrix}AB&0\\B&0\end{pmatrix}\begin{pmatrix}I_m&A\\0&I_n\end{pmatrix}=\begin{pmatrix}AB&ABA\\B&BA\end{pmatrix}=\begin{pmatrix}I_m&A\\0&I_n\end{pmatrix}\begin{pmatrix}0&0\\B&BA\end{pmatrix}.$$So the two block matrices on the outside are similar, and by Theorem 3.16(4) their characteristic polynomials are $p_{AB}(x)\,x^n$ and $x^m\,p_{BA}(x)$. (4) Apply (3) with the $n\times1$ matrix $u$ and the $1\times n$ matrix $v^{\mathsf T}$: $v^{\mathsf T}u$ is $1\times1$ with characteristic polynomial $x-v^{\mathsf T}u$, so $x\,p_{uv^{\mathsf T}}(x)=x^n(x-v^{\mathsf T}u)$. For $J=\mathbf{1}\mathbf{1}^{\mathsf T}$ the eigenvalues are $n$ and $0$ ($n-1$ times), and $aI+bJ$ has eigenvalues $a+b\mu$ for each eigenvalue $\mu$ of $J$, by (2). ∎
Cayley–Hamilton and the Minimal Polynomial
- $m_A$ exists and is unique, and for every $f\in\mathbb{F}[x]$: $f(A)=0\iff m_A\mid f$.
- Similar matrices have the same minimal polynomial.
- For a block diagonal matrix, $m_{\operatorname{diag}(A_1,\dots,A_k)}=\operatorname{lcm}(m_{A_1},\dots,m_{A_k})$.
Proof. (1) The $n^2+1$ matrices $I,A,\dots,A^{n^2}$ lie in the $n^2$-dimensional space $M_n(\mathbb{F})$, so they are dependent: some nonzero polynomial annihilates $A$, and there is one of least degree, which we scale to be monic. If $f(A)=0$, divide: $f=qm_A+r$ with $\deg r\lt\deg m_A$; then $r(A)=f(A)-q(A)m_A(A)=0$, and minimality forces $r=0$. Two monic annihilating polynomials of least degree divide each other, so they are equal. (2) $f(P^{-1}AP)=P^{-1}f(A)P$, so the same polynomials annihilate both. (3) $f(\operatorname{diag}(A_i))=\operatorname{diag}(f(A_i))$ vanishes iff every $m_{A_i}$ divides $f$, iff the least common multiple divides $f$. ∎
Proof. Let $B(x)=\operatorname{adj}(xI-A)$, the adjugate (transposed cofactor matrix). Its entries are $(n-1)\times(n-1)$ minors of $xI-A$, polynomials of degree at most $n-1$, so $B(x)=B_0+B_1x+\cdots+B_{n-1}x^{n-1}$ with constant matrices $B_k$. The adjugate identity $M\operatorname{adj}M=(\det M)I$ gives
$$(xI-A)\,B(x)=p_A(x)\,I,\qquad p_A(x)=x^n+c_{n-1}x^{n-1}+\cdots+c_0 .$$Comparing coefficients of powers of $x$: $-AB_0=c_0I$; $B_{k-1}-AB_k=c_kI$ for $1\le k\le n-1$; and $B_{n-1}=I$. Multiply the equation for $x^k$ on the left by $A^k$ and add all of them:
$$p_A(A)=-AB_0+\sum_{k=1}^{n-1}\left(A^kB_{k-1}-A^{k+1}B_k\right)+A^nB_{n-1}.$$The sum telescopes: every term $A^{k+1}B_k$ appears once with each sign. So $p_A(A)=0$, and $m_A\mid p_A$ by Theorem 3.19(1). ∎
Note what the proof does not do: "substitute $x=A$ in $\det(xI-A)$ to get $\det(A-A)=0$" is not an argument — $p_A(A)$ is a matrix, while $\det(AI-A)$ would be a number.
- $\lambda\in\mathbb{F}$ is an eigenvalue of $A$ if and only if $m_A(\lambda)=0$. Over $\mathbb{C}$, if $p_A=\prod_i(x-\lambda_i)^{a_i}$ with distinct $\lambda_i$, then $m_A=\prod_i(x-\lambda_i)^{k_i}$ with $1\le k_i\le a_i$.
- $A$ is invertible iff $m_A(0)\ne0$ iff $p_A(0)\ne0$; then $A^{-1}$ is a polynomial in $A$: if $p_A(x)=x\,q(x)+c_0$ with $c_0\ne0$, then $A^{-1}=-c_0^{-1}q(A)$.
Proof. (1) If $Av=\lambda v$ with $v\ne0$, then $0=m_A(A)v=m_A(\lambda)v$, so $m_A(\lambda)=0$. Conversely every root of $m_A$ is a root of $p_A$ because $m_A\mid p_A$. Over $\mathbb{C}$ this says $m_A$ contains each factor $x-\lambda_i$ at least once and, dividing $p_A$, at most $a_i$ times. (2) $0$ is an eigenvalue iff it is a root of $m_A$ iff of $p_A$. From $p_A(A)=0$: $A\,q(A)=-c_0I$. ∎
- Factor $p_A=\prod(x-\lambda_i)^{a_i}$. The candidates are $\prod(x-\lambda_i)^{k_i}$ with $1\le k_i\le a_i$; test them in order of increasing degree.
- For block diagonal matrices take the lcm of the blocks' minimal polynomials; a Jordan block $J_k(\lambda)$ has minimal polynomial $(x-\lambda)^k$.
- Faster, with ranks: the exponent $k_i$ of $(x-\lambda_i)$ in $m_A$ is the least $k$ with $\operatorname{rank}(A-\lambda_iI)^k=\operatorname{rank}(A-\lambda_iI)^{k+1}$ (Theorem 3.25).
- If $A$ satisfies some identity $f(A)=0$, then $m_A$ is one of the monic divisors of $f$ — exclude the ones that do not annihilate $A$.
Diagonalisability
- $A$ is diagonalisable over $\mathbb{F}$;
- $\mathbb{F}^n$ has a basis of eigenvectors of $A$ — and then $P^{-1}AP=D$ for the matrix $P$ whose columns are these eigenvectors, with the corresponding eigenvalues on the diagonal of $D$;
- $\sum_\lambda g(\lambda)=n$; equivalently, $p_A$ splits over $\mathbb{F}$ and $g(\lambda)=a(\lambda)$ for every eigenvalue;
- $m_A$ splits over $\mathbb{F}$ into distinct linear factors: $m_A=(x-\lambda_1)\cdots(x-\lambda_k)$ with $\lambda_1,\dots,\lambda_k$ distinct.
Proof. (1)$\iff$(2): $AP=PD$ says exactly that the $j$-th column $p_j$ of $P$ satisfies $Ap_j=d_jp_j$, and $P$ is invertible iff its columns form a basis. (2)$\Rightarrow$(3): grouping a basis of eigenvectors by eigenvalue gives independent vectors in each $E_\lambda$, so $\sum g(\lambda)\ge n$; the reverse inequality is Theorem 3.17(2). (3)$\Rightarrow$(2): by Theorem 3.17(1) the union of bases of the eigenspaces is independent, and it has $\sum g(\lambda)=n$ vectors. The two forms of (3) are equivalent because $\sum g\le\sum a\le n$ with $g(\lambda)\le a(\lambda)$ termwise, and $\sum a=n$ iff $p_A$ splits.
(1)$\Rightarrow$(4): if $D$ has distinct diagonal values $\lambda_1,\dots,\lambda_k$, then $\prod_i(D-\lambda_iI)=0$ (each diagonal position meets one zero factor), so $\prod_i(x-\lambda_i)$ annihilates $A=PDP^{-1}$; $m_A$ divides it and has every $\lambda_i$ as a root (Theorem 3.21), so it equals it. (4)$\Rightarrow$(2): consider the Lagrange polynomials
$$L_i(x)=\prod_{j\ne i}\frac{x-\lambda_j}{\lambda_i-\lambda_j},\qquad i=1,\dots,k.$$The polynomial $L_1+\cdots+L_k-1$ has degree $\lt k$ and vanishes at the $k$ points $\lambda_1,\dots,\lambda_k$, so $L_1+\cdots+L_k=1$. Hence every $v\in\mathbb{F}^n$ satisfies $v=\sum_iL_i(A)v$. Moreover $(x-\lambda_i)L_i(x)$ is a constant multiple of $m_A$, so $(A-\lambda_iI)L_i(A)v=0$: each $L_i(A)v$ lies in $E_{\lambda_i}$. Thus the eigenvectors span $\mathbb{F}^n$, and a basis of eigenvectors can be extracted from them. ∎
- If $A$ has $n$ distinct eigenvalues in $\mathbb{F}$, it is diagonalisable over $\mathbb{F}$.
- If $A\in M_n(\mathbb{C})$ satisfies $A^k=I$ for some $k\ge1$, it is diagonalisable over $\mathbb{C}$ with eigenvalues $k$-th roots of unity. If $A^2=A$ (idempotent), it is diagonalisable with eigenvalues in $\{0,1\}$, and $\operatorname{rank}A=\operatorname{tr}A$. If $A^2=I$, it is diagonalisable over $\mathbb{R}$ (when $A$ is real) with eigenvalues $\pm1$.
- A diagonalisable matrix with a single eigenvalue $\lambda$ is $\lambda I$. In particular a nilpotent matrix ($N^k=0$ for some $k$) is diagonalisable only if $N=0$.
- If $A=PDP^{-1}$, then $f(A)=Pf(D)P^{-1}$ for every polynomial $f$; in particular $A^m=PD^mP^{-1}$, and $f(A)$ is diagonalisable.
- Real symmetric matrices, Hermitian matrices and, more generally, normal matrices are diagonalisable — even by an orthogonal or unitary $P$ (§3.3).
Proof. (1) Each $g(\lambda)\ge1$, so $\sum g\ge n$. (2) $m_A$ divides $x^k-1$, whose roots $e^{2\pi ij/k}$ are distinct; it divides $x(x-1)$, resp. $(x-1)(x+1)$, whose roots are distinct and real; apply Theorem 3.22(4). For an idempotent, $D=\operatorname{diag}(1,\dots,1,0,\dots,0)$, so both rank and trace count the $1$'s. (3) $D=\lambda I$ gives $A=P(\lambda I)P^{-1}=\lambda I$; a nilpotent matrix has only the eigenvalue $0$ (Theorem 3.18(2): $\lambda^k=0$). (4) $(PDP^{-1})^m=PD^mP^{-1}$, then add up. (5) is proved in §3.3. ∎
The Jordan Canonical Form from Rank Data
The Jordan block of size $k$ for $\lambda$ is the $k\times k$ matrix $J_k(\lambda)$ with $\lambda$ on the diagonal, $1$ on the superdiagonal and $0$ elsewhere; $J_1(\lambda)=(\lambda)$. For instance
$$J_3(\lambda)=\begin{pmatrix}\lambda&1&0\\0&\lambda&1\\0&0&\lambda\end{pmatrix}.$$A Jordan matrix is a block diagonal matrix $J_{k_1}(\lambda_1)\oplus\cdots\oplus J_{k_s}(\lambda_s)$ (the $\lambda_i$ need not be distinct). A partition of $a$ is a way of writing $a$ as a sum of positive integers, ignoring order; $p(a)$ denotes the number of partitions of $a$, and $P(a,m)$ the number whose largest part is exactly $m$. Thus $p(1),\dots,p(7)=1,2,3,5,7,11,15$.
Proof (sketch; full proof in the 📚 box). Step 1: split the space. If $p_A=\prod_i(x-\lambda_i)^{a_i}$, the factors $(x-\lambda_i)^{a_i}$ are pairwise coprime, and a Lagrange/Bézout argument like the one in Theorem 3.22 shows $\mathbb{C}^n=\bigoplus_iK_i$ with $K_i=N\left((A-\lambda_iI)^{a_i}\right)$, the generalised eigenspaces. Each $K_i$ is mapped into itself by $A$, has dimension $a_i$, and $N_i=(A-\lambda_iI)|_{K_i}$ is nilpotent. Step 2: nilpotent maps. A nilpotent operator $N$ has a basis consisting of chains $v,Nv,\dots,N^{m-1}v$ with $N^mv=0$; in the reversed chain $N^{m-1}v,\dots,Nv,v$ the matrix of $N$ is $J_m(0)$. Such a basis is built from the top: choose vectors whose images under $N^{m-1}$ form a basis of $\operatorname{im}N^{m-1}$, extend downwards level by level. So $A|_{K_i}=\lambda_iI+N_i$ is a direct sum of blocks $J_m(\lambda_i)$. Step 3: uniqueness. The number of blocks of each size is computed from ranks in Theorem 3.25, and ranks are similarity invariants. ∎
- the number of blocks for $\lambda$ is $g(\lambda)=n-r_1$, and their sizes add up to $a(\lambda)$;
- the number of blocks for $\lambda$ of size $\ge k$ is $r_{k-1}-r_k$, and of size exactly $k$ is $r_{k-1}-2r_k+r_{k+1}$;
- the largest block for $\lambda$ has size equal to the exponent of $(x-\lambda)$ in $m_A$, which is the least $k$ with $r_k=r_{k+1}$ (and then $r_k=n-a(\lambda)$);
- the number of linearly independent eigenvectors of $A$ is the total number of blocks; $A$ is diagonalisable iff every block is $1\times1$, iff $r_1=r_2$ for every eigenvalue.
Proof. Ranks of block diagonal matrices add, and conjugation does not change ranks, so we may compute block by block in $J_A$. For a block $J_m(\mu)$ with $\mu\ne\lambda$, $J_m(\mu)-\lambda I$ is triangular with nonzero diagonal, hence invertible: it contributes $m$ to every $r_k$. For a block $J_m(\lambda)$, $J_m(\lambda)-\lambda I=J_m(0)$ shifts the standard basis, $e_1\mapsto0$, $e_j\mapsto e_{j-1}$; so $J_m(0)^k$ has $1$'s on the $k$-th superdiagonal and
$$\operatorname{rank}J_m(0)^k=\max(m-k,0).$$Therefore, summing over the blocks $J_m(\lambda)$ for the eigenvalue $\lambda$,
$$r_{k-1}-r_k=\sum_{J_m(\lambda)}\Big[\max(m-k+1,0)-\max(m-k,0)\Big],$$and each bracket is $1$ if $m\ge k$ and $0$ otherwise: this is (2), and (1) is the case $k=1$ together with Theorem 3.17. The block $J_m(\lambda)$ has minimal polynomial $(x-\lambda)^m$, so by Theorem 3.19(3) the exponent of $(x-\lambda)$ in $m_A$ is the largest $m$; and $r_k$ stops decreasing exactly when $k$ reaches it. (4) Each block contributes exactly one eigenvector ($e_1$ of the block), so $\sum_\lambda g(\lambda)$ is the number of blocks; all blocks have size $1$ iff no block has size $\ge2$ iff $r_1-r_2=0$ for every $\lambda$. ∎
It helps to picture the blocks for $\lambda$ as columns of dots (a dot diagram), one column of height $m$ for each block $J_m(\lambda)$, aligned at the bottom. Then $r_{k-1}-r_k$ is the number of dots in row $k$ from the bottom: the bottom row counts the blocks, i.e. the eigenvectors. Interactive 2 animates this.
- if every $a(\lambda)\le3$, the matrix is determined up to similarity by $p_A$ and $m_A$;
- if every $a(\lambda)\le6$, it is determined by $p_A$, $m_A$ and the geometric multiplicities $g(\lambda)$;
- both bounds are sharp: $J_2(0)\oplus J_2(0)$ and $J_2(0)\oplus J_1(0)\oplus J_1(0)$ share $p=x^4$ and $m=x^2$; $J_3(0)\oplus J_3(0)\oplus J_1(0)$ and $J_3(0)\oplus J_2(0)\oplus J_2(0)$ share $p=x^7$, $m=x^3$ and $g=3$.
Proof. The first sentence is Theorem 3.25. For $a\le3$ the partitions are $(1)$; $(2),(1,1)$; $(3),(2,1),(1,1,1)$ — different largest parts in each case. For $a\le6$ one lists the $1+2+3+5+7+11=29$ partitions and checks that no two partitions of the same $a$ have both the same largest part and the same number of parts; the first coincidence occurs at $a=7$ with $(3,3,1)$ and $(3,2,2)$ (and at $a=4$ for largest part alone: $(2,2)$, $(2,1,1)$). The examples in (3) realise these coincidences; they are not similar because, for instance, $\operatorname{rank}N^2$ is $2$ for $(3,3,1)$ and $1$ for $(3,2,2)$. ∎
Proof. By Theorem 3.24 similarity classes correspond to Jordan matrices up to block order, i.e. to a choice, independently for each $\lambda_i$, of a partition of $a_i$ (the block sizes). Prescribing $m_A$ prescribes the largest part $m_i$ (Theorem 3.25(3)). A nilpotent matrix has $p_A=x^n$, and $N^k=0$ means that every block has size at most $k$. ∎
- Every $A\in M_n(\mathbb{C})$ is similar to its transpose.
- If $N\in M_n(\mathbb{C})$ is nilpotent then $N^n=0$; its index (least $k$ with $N^k=0$) is the size of its largest Jordan block, and the index is $n$ iff $\operatorname{rank}N=n-1$.
- $J_k(\lambda)^m=\sum_{j=0}^{k-1}\binom mj\lambda^{m-j}J_k(0)^j$: the $j$-th superdiagonal of $J_k(\lambda)^m$ is constant, equal to $\binom mj\lambda^{m-j}$. Consequently $A^m\to0$ as $m\to\infty$ if and only if every eigenvalue of $A$ satisfies $|\lambda|\lt1$.
Proof. (1) Let $W$ be the $k\times k$ matrix with $1$'s on the antidiagonal; reversing the order of the basis, $WJ_k(\lambda)W=J_k(\lambda)^{\mathsf T}$ and $W=W^{-1}$. Doing this blockwise shows $J_A\sim J_A^{\mathsf T}$, and $A\sim J_A$, $A^{\mathsf T}\sim J_A^{\mathsf T}$. (2) All blocks of $N$ are $J_m(0)$ with $m\le n$, and $J_m(0)^k=0$ iff $k\ge m$. A single block of size $n$ is the same as $g(0)=n-\operatorname{rank}N=1$. (3) $J_k(\lambda)=\lambda I+J_k(0)$, the two terms commute, and $J_k(0)^j=0$ for $j\ge k$; apply the binomial theorem. If $|\lambda|\lt1$, each entry $\binom mj\lambda^{m-j}$ is a polynomial in $m$ times $|\lambda|^m$, which tends to $0$; if $|\lambda|\ge1$ the diagonal entries $\lambda^m$ do not tend to $0$. Apply this to $A^m=PJ_A^mP^{-1}$. ∎
- Factor $p_A$; for each eigenvalue note $a(\lambda)$ — the total size of its blocks.
- Compute $g(\lambda)=n-\operatorname{rank}(A-\lambda I)$ — the number of blocks. If $g=a$, all blocks are $1\times1$.
- If the data are still ambiguous, compute $\operatorname{rank}(A-\lambda I)^2,\dots$ and use $r_{k-1}-r_k=\#\{\text{blocks of size}\ge k\}$; or use the exponent in $m_A$ as the largest block.
- Check: the sizes for each $\lambda$ add up to $a(\lambda)$, the number of blocks is $g(\lambda)$, and the largest block matches $m_A$.
- Eigenvalues are not additive or multiplicative. $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ and $B=\begin{pmatrix}0&0\\1&0\end{pmatrix}$ have only the eigenvalue $0$, but $A+B$ has eigenvalues $\pm1$ and $AB=\operatorname{diag}(1,0)$ has eigenvalue $1$. (For commuting matrices the eigenvalues of $A+B$ are sums $\lambda_i+\mu_i$ in a suitable order.)
- The same characteristic polynomial does not give similarity: $I_2$ and $J_2(1)$. The same $p_A$ and $m_A$ do not either, from size $4$ on: $J_2(0)\oplus J_2(0)$ and $J_2(0)\oplus J_1(0)\oplus J_1(0)$. Nor do $p_A$, $m_A$ and all $g(\lambda)$, from size $7$ on: $(3,3,1)$ versus $(3,2,2)$.
- A repeated eigenvalue does not prevent diagonalisability ($I$, or matrix $A$ of Example 1); distinct eigenvalues are sufficient, not necessary.
- $AB$ and $BA$ have the same eigenvalues but need not be similar. $A=\operatorname{diag}(1,0)$, $B=E_{12}$: $AB=E_{12}\ne0$ while $BA=0$.
- Row operations change eigenvalues. Never row-reduce $A$ before computing $p_A$; row-reduce $A-\lambda I$ only to find its rank or null space.
- $A$ and $A^{\mathsf T}$ have the same eigenvalues but usually different eigenvectors: $J_2(0)$ has eigenvector $e_1$, its transpose $e_2$.
- The field matters. The rotation $\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ has no real eigenvalues and is diagonalisable over $\mathbb{C}$ but not over $\mathbb{R}$; it satisfies $A^4=I$, so "$A^k=I\Rightarrow$ diagonalisable" is a statement over $\mathbb{C}$. Over $\mathbb{F}_2$, $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ satisfies $A^2=I$ and is not diagonalisable, because $x^2-1=(x-1)^2$ there.
- $m_A=p_A$ does not mean distinct eigenvalues ($J_2(0)$), and $\deg m_A\lt n$ does not mean diagonalisable ($J_2(0)\oplus(0)$ has $m=x^2$, $n=3$).
- $A^2$ diagonalisable does not make $A$ diagonalisable ($J_2(0)$), unless $A$ is invertible.
- Geometric multiplicity is $n-\operatorname{rank}(A-\lambda I)$, not the rank. And it counts eigenvectors (blocks), not the size of the eigenvalue's blocks.
- Cayley–Hamilton is not "$\det(A-A)=0$". See the remark after Theorem 3.20.
- Nilpotent $\ne$ zero eigenvalues over $\mathbb{R}$. A real matrix all of whose real eigenvalues are $0$, such as $(0)\oplus\begin{pmatrix}0&-1\\1&0\end{pmatrix}$, need not be nilpotent; over $\mathbb{C}$, "all eigenvalues $0$" does imply $N^n=0$ (Cayley–Hamilton).
- CSIR Unit 1 · §9.1 Eigenvalues, the Minimal Polynomial and Cayley–Hamilton — eigenvalues of operators, the principal-minor expansion of $p_A$ in detail, the division property of $m_T$, cyclic vectors and companion matrices, and spectral mapping with proofs.
- CSIR Unit 1 · §9.2 Diagonal and Triangular Forms — invariant subspaces, the primary decomposition theorem with full proof, the minimal-polynomial criterion, Schur's unitary triangularisation, and simultaneous diagonalisation of commuting operators.
- CSIR Unit 1 · §9.3 The Jordan Canonical Form — generalised eigenspaces and the Fitting decomposition, the complete existence and uniqueness proof via nilpotent chains, the Jordan–Chevalley decomposition, the table of $P(a,m)$, and the rational canonical form over any field.
Two Labs
Exercises
- 1 Find the eigenvalues and eigenspaces of $\begin{pmatrix}1&2\\3&2\end{pmatrix}$.
- 2 Find the characteristic polynomial and the eigenvalues of $\begin{pmatrix}0&0&6\\1&0&-11\\0&1&6\end{pmatrix}$.
- 3 $A\in M_3(\mathbb{R})$ has $\operatorname{tr}A=6$, $\operatorname{tr}A^2=14$ and $\det A=6$. Find its eigenvalues.
- 4 Show that a nilpotent matrix has $p_A=x^n$. Conversely, if $A\in M_n(\mathbb{C})$ has only the eigenvalue $0$, show that $A$ is nilpotent.
- 5 Every row of $A\in M_n(\mathbb{R})$ sums to $s$. Show that $s$ is an eigenvalue of $A$, and that if $A$ is invertible every row of $A^{-1}$ sums to $1/s$.
- 6 Find the eigenvalues and the determinant of the $n\times n$ matrix with $0$ on the diagonal and $1$ everywhere else.
- 7 Show that $AB$ and $BA$ need not be similar, using $A=\operatorname{diag}(1,0)$ and $B=\begin{pmatrix}0&1\\0&0\end{pmatrix}$.
- 8 For which real $k$ does $\begin{pmatrix}1&k\\1&1\end{pmatrix}$ have real eigenvalues? Distinct real eigenvalues? Is it diagonalisable when $k=0$?
- 9 $A\in M_3(\mathbb{R})$ satisfies $\det(A-I)=\det(A-2I)=\det(A+I)=0$. Find $\operatorname{tr}A$, $\det A$ and $\det(A^2+2I)$.
- 1 Verify the Cayley–Hamilton theorem for $A=\begin{pmatrix}1&2\\3&4\end{pmatrix}$ and use it to write $A^{-1}$ as a polynomial in $A$.
- 2 $A\in M_3(\mathbb{R})$ has $p_A(x)=x^3-2x^2-x+2$. Write $A^4$ and $A^{-1}$ in the form $\alpha A^2+\beta A+\gamma I$.
- 3 Find the minimal polynomial of $J_2(2)\oplus J_1(2)\oplus J_1(5)$.
- 4 Find the minimal polynomial of the $3\times3$ matrix all of whose entries are $1$.
- 5 A real matrix $A$ has $m_A(x)=x^2+1$. Show that its size $n$ is even and that $A$ is diagonalisable over $\mathbb{C}$ but not over $\mathbb{R}$.
- 6 $A\in M_n(\mathbb{C})$, $A\ne0$, satisfies $A^3=A^2$. List the possible minimal polynomials, and show that $A^2$ is idempotent.
- 7 Prove that the inverse of an invertible matrix is a polynomial in the matrix. Is the polynomial unique?
- 8 $A\in M_3(\mathbb{C})$ has $m_A=(x-1)^2$. Find $p_A$ and the Jordan form of $A$.
- 1 Diagonalise $A=\begin{pmatrix}4&-2\\1&1\end{pmatrix}$ and find a formula for $A^n$.
- 2 Is $\begin{pmatrix}2&1&0\\0&2&0\\0&0&3\end{pmatrix}$ diagonalisable?
- 3 For which real $a,b$ is $\begin{pmatrix}1&a\\0&b\end{pmatrix}$ diagonalisable?
- 4 Prove that every idempotent matrix is diagonalisable and that its rank equals its trace.
- 5 $A\in M_n(\mathbb{R})$ satisfies $A^2=I$. Show that $A$ is diagonalisable over $\mathbb{R}$ and that $\operatorname{tr}A\equiv n\pmod 2$.
- 6 Give a real $2\times2$ matrix that is diagonalisable over $\mathbb{C}$ but not over $\mathbb{R}$.
- 7 Prove that a nonzero nilpotent matrix is not diagonalisable.
- 8 Show that if $A$ is diagonalisable then so is $f(A)$ for every polynomial $f$. Is the converse true?
- 9 Modelling. A city bike-share scheme has two stations. Each day $80\%$ of the bikes at station $A$ stay there and $20\%$ are ridden to $B$; at $B$, $70\%$ stay and $30\%$ go to $A$. With $x_n=(a_n,b_n)^{\mathsf T}$ the fractions of the fleet at $A$ and $B$ on day $n$, $x_{n+1}=Mx_n$. Find $M$, diagonalise it, and find the long-run distribution of the bikes and how fast it is approached.
- 1 Find the Jordan form and the minimal polynomial of $\begin{pmatrix}3&1&0\\-1&1&0\\0&0&2\end{pmatrix}$.
- 2 A nilpotent $N\in M_6(\mathbb{C})$ has $\operatorname{rank}N=3$, $\operatorname{rank}N^2=1$, $N^3=0$. Find its Jordan form.
- 3 List all possible Jordan forms of $A\in M_5(\mathbb{C})$ with $p_A=(x-2)^3(x-5)^2$ and $m_A=(x-2)^2(x-5)$.
- 4 Show that there are exactly $5$ similarity classes of $4\times4$ complex nilpotent matrices, and give the rank sequence $(\operatorname{rank}N,\operatorname{rank}N^2,\operatorname{rank}N^3)$ of each.
- 5 How many similarity classes are there of $6\times6$ complex matrices with $p_A=(x-1)^4(x+1)^2$?
- 6 Prove that $J_k(\lambda)$ is similar to its transpose.
- 7 Suppose $p_A=m_A$ for $A\in M_n(\mathbb{C})$. Show that every eigenvalue has exactly one Jordan block, so $g(\lambda)=1$ for every eigenvalue.
- 8 Compute $J_3(2)^5$.
E1. GATE MA · MCQ Let $A$ be the $5\times5$ real matrix with every diagonal entry equal to $3$ and every off-diagonal entry equal to $1$. Then $\det A$ is (A) $112$; (B) $7$; (C) $48$; (D) $243$.
E2. GATE MA · MCQ The minimal polynomial of $A=\begin{pmatrix}1&1&1\\0&1&1\\0&0&1\end{pmatrix}$ is (A) $x-1$; (B) $(x-1)^2$; (C) $(x-1)^3$; (D) $x^3-1$.
E3. GATE MA · MSQ Let $A\in M_3(\mathbb{R})$ satisfy $A^3=A$, $\operatorname{tr}A=1$ and $\det A=0$. Which of the following is/are true? (One or more options may be correct.) (A) $A$ is diagonalisable over $\mathbb{R}$. (B) $\operatorname{rank}A=2$. (C) $A^2=A$. (D) $\det(A+I)=2$.
E4. GATE MA · MSQ Let $N\in M_4(\mathbb{C})$ be nilpotent with $\operatorname{rank}N=2$. Which of the following is/are necessarily true? (One or more options may be correct.) (A) $N^2=0$. (B) $N^3=0$. (C) If $N^2=0$, then $N$ is similar to $J_2(0)\oplus J_2(0)$. (D) $\operatorname{rank}N^2\le1$.
E5. GATE MA · NAT The number of similarity classes of $8\times8$ complex matrices $A$ with $A^3=0$ is ______.
E6. GATE MA · NAT $A\in M_3(\mathbb{R})$ has characteristic polynomial $x^3-7x^2+14x-8$. The value of $\operatorname{tr}(A^{-1})$ is ______ (correct to two decimal places).
Suggested plan — 3 lectures × 60 min.
- Eigenvalues. Definitions 3.9–3.10, Theorems 3.15–3.17 with proofs (the block form $\begin{pmatrix}\lambda I_g&C\\0&D\end{pmatrix}$ is worth drawing). Example 1 on the board — let half the class do $A$, half do $B$, then compare. Project Interactive 1 on the shear and the rotation. Theorem 3.18 and Example 2 as a quick-fire GATE drill.
- Polynomials in a matrix. Definition 3.11, Theorem 3.19, the adjugate proof of Cayley–Hamilton (Theorem 3.20) and the warning after it, Theorem 3.21, Examples 3 and 4. Then diagonalisability: Theorem 3.22 with the Lagrange-polynomial proof, Theorem 3.23 and Example 5 as a prove-or-disprove discussion.
- Jordan form. Definition 3.13, Theorem 3.24 (sketch), Theorem 3.25 with the rank formula $\operatorname{rank}J_m(0)^k=\max(m-k,0)$. Project Interactive 2 and step through $(3,3,1)$ and $(3,2,2)$. Theorems 3.26–3.28, the strategy box, Examples 6 and 7. Finish with GATE corner E4 and E5.
Discussion questions. (1) Why can the characteristic polynomial never detect whether a matrix is diagonalisable, while the minimal polynomial always can? (2) A $3\times3$ matrix has a single eigenvalue. How many different Jordan forms can it have, and which single number (a rank) decides between them? (3) In Interactive 1, the cross product $c(\theta)$ is a quadratic form in $v$. Why does a defective matrix produce a double zero, and what would a triple root of $p_A$ look like for a $3\times3$ matrix?
Misconceptions to watch for. Row-reducing $A$ and reading eigenvalues off the result; treating $g(\lambda)$ as $\operatorname{rank}(A-\lambda I)$; "repeated eigenvalue $\Rightarrow$ not diagonalisable"; "$p_A$ and $m_A$ determine the matrix up to similarity" (false from $n=4$); the fake proof of Cayley–Hamilton; forgetting that diagonalisability depends on the field; adding eigenvalues of non-commuting matrices.
Suggested homework. I(2, 3, 5, 6, 9), II(1, 2, 5, 6), III(1, 3, 5, 8, 9), IV(1, 2, 3, 5, 8); GATE corner E1, E3, E6.
Research question. Will an age-structured animal population grow or decline in the long run, at what rate, and with what age profile — and why are the answers an eigenvalue and an eigenvector of a $3\times3$ matrix?
Why it matters. Wildlife agencies, fisheries scientists and conservation biologists (for sea turtles, seabirds, fish stocks) use Leslie matrix models to decide which life stage to protect. If $x_t$ lists the numbers of juveniles, sub-adults and adults in year $t$, then $x_{t+1}=Lx_t$, so $x_t=L^tx_0$, and the long-run behaviour of $L^t$ is governed by the eigenvalue of largest modulus (Theorem 3.28(3)).
What to do.
- Take $L=\begin{pmatrix}0&1.5&1.2\\0.6&0&0\\0&0.5&0\end{pmatrix}$: sub-adults and adults have $1.5$ and $1.2$ offspring per year, $60\%$ of juveniles and $50\%$ of sub-adults survive to the next stage. Show that $p_L(x)=x^3-0.9x-0.36$.
- Show that $p_L(1)\lt0$ and conclude that $L$ has a real eigenvalue $\lambda_1\gt1$. Find it numerically (Newton's method or a root finder) and the corresponding eigenvector, normalised to sum $1$ — the stable age distribution.
- Start from $x_0=(100,0,0)$ and iterate $x_{t+1}=Lx_t$ for $30$ years (the power method). Plot the ratios of total population in successive years and the proportions in each class. How fast do they settle, and how is that speed related to the other two eigenvalues?
- The net reproductive number is $R_0=1.5\cdot0.6+1.2\cdot0.6\cdot0.5$. Show that $\lambda_1\gt1$ exactly when $R_0\gt1$ (divide $p_L(\lambda)=0$ by $\lambda^3$). Which survival rate would you try to improve to make the population grow fastest?
Tools: Python with numpy (numpy.linalg.eig, a loop for the power method) or a spreadsheet.
Hand in: the characteristic polynomial and its roots; the stable age distribution; the plots of step 3; a short paragraph answering the conservation question of step 4 with numbers (change one survival rate by $0.05$ and recompute $\lambda_1$).
Going further. Read about the Perron–Frobenius theorem, which guarantees a positive dominant eigenvalue with a positive eigenvector for matrices like $L$, and about elasticity analysis, which measures how $\lambda_1$ responds to each entry of $L$ via the left and right eigenvectors.
🧪 Self-check quiz
Q1. $A\in M_3(\mathbb{R})$ has eigenvalues $1,2,3$. Then $\det(A+I)$ equals
Q2. Which of these matrices is not diagonalisable over $\mathbb{C}$?
Q3. The minimal polynomial of the Jordan matrix $J_2(0)\oplus J_1(0)\oplus J_3(0)$ is
Q4. In the Jordan form of $A\in M_n(\mathbb{C})$, the number of Jordan blocks for the eigenvalue $\lambda$ equals
Q5. How many similarity classes of $3\times3$ complex nilpotent matrices are there?