9.1 Eigenvalues, the Minimal Polynomial and Cayley–Hamilton
An eigenvector is a direction that a linear map only stretches. Eigenvalues are the roots of the characteristic polynomial, every square matrix satisfies its own characteristic polynomial, and the minimal polynomial records exactly which polynomial identities a matrix obeys.
A linear operator on a space of dimension $n$ is described by $n^2$ numbers, and most of them depend on the basis we happen to use (§8.2). Some information does not: the directions that the operator merely stretches, and the stretching factors. If $Av=\lambda v$, then $A^kv=\lambda^kv$, $e^{tA}v=e^{\lambda t}v$, and on the line through $v$ the whole operator acts as multiplication by one number. This is why eigenvalues decide the long-run behaviour of iterated maps (population models, Markov chains, the Fibonacci recursion, the power method behind web search), the stability of $x'=Ax$, the natural frequencies of vibrating structures, and — in the next two sections — how simple a matrix can be made by a change of basis.
This section builds the toolkit: eigenvalues and eigenspaces; the characteristic polynomial and what its coefficients say (trace, determinant, principal minors); algebraic and geometric multiplicity; polynomials in an operator; the Cayley–Hamilton theorem, proved in full; the minimal polynomial; eigenvalues of $f(A)$; the surprising relation between $AB$ and $BA$; and companion matrices, which show that every monic polynomial is a characteristic polynomial. In CSIR NET this is one of the most heavily examined topics of Unit 1: Part B asks for $\det$, $\operatorname{tr}$ or the eigenvalues of a matrix built from another one, and Part C asks which statements about characteristic and minimal polynomials are true.
Throughout, $V$ is a vector space over a field $\mathbb{F}$ with $\dim V=n\ge1$, $T\in\mathcal{L}(V)$, and $A\in M_n(\mathbb{F})$; as in Chapter 8, a matrix $A$ is identified with the operator $x\mapsto Ax$ on $\mathbb{F}^n$.
Eigenvalues, Eigenvectors and Eigenspaces
A scalar $\lambda\in\mathbb{F}$ is an eigenvalue of $T$ if there is a nonzero vector $v\in V$ with $Tv=\lambda v$; every such $v$ is an eigenvector of $T$ belonging to $\lambda$. The subspace
$$E_\lambda=E_\lambda(T)=\ker(T-\lambda I)=\{v\in V:Tv=\lambda v\}$$is the eigenspace of $\lambda$; it consists of the eigenvectors together with $0$. The set $\sigma(T)$ of all eigenvalues is the spectrum of $T$. For $A\in M_n(\mathbb{F})$ these notions refer to $x\mapsto Ax$ on $\mathbb{F}^n$. Note that eigenvalues are required to lie in $\mathbb{F}$: the same real matrix may have no eigenvalue over $\mathbb{R}$ and two over $\mathbb{C}$.
Proof. (a)$\iff$(b): $Tv=\lambda v$ with $v\ne0$ says exactly that $0\ne v\in\ker(T-\lambda I)$. (b)$\iff$(c)$\iff$(d): for an operator on a finite-dimensional space, rank–nullity (§8.2) gives $\dim\ker S+\dim\operatorname{im}S=n$, so $S$ is injective iff $\ker S=0$ iff $\operatorname{im}S=V$ iff $S$ is surjective, and then $S$ is bijective. (d)$\iff$(e): $[T-\lambda I]_{\mathcal{B}}=A-\lambda I$, and an operator is invertible iff its matrix is, iff the determinant of the matrix is nonzero (§8.3); $\det(\lambda I-A)=(-1)^n\det(A-\lambda I)$. A different basis replaces $A$ by $P^{-1}AP$ and $$\det(\lambda I-P^{-1}AP)=\det\big(P^{-1}(\lambda I-A)P\big)=\det(\lambda I-A),$$ so (e) does not depend on $\mathcal{B}$. $E_\lambda$ is a kernel, hence a subspace; if $Tv=\lambda v$ then $T(Tv)=\lambda(Tv)$, so $Tv\in E_\lambda$. The dimension formula is rank–nullity for $A-\lambda I$. The last statement is the case $\lambda=0$. ∎
Finite dimension is essential for (b)$\iff$(c). On the space of real sequences the right shift $R(x_1,x_2,\dots)=(0,x_1,x_2,\dots)$ is injective but not surjective, and it has no eigenvalues: $Rx=\lambda x$ gives $\lambda x_1=0$, $\lambda x_{k+1}=x_k$, which forces $x=0$ both for $\lambda=0$ and for $\lambda\ne0$. At the other extreme, the differentiation operator $D$ on $C^\infty(\mathbb{R})$ has every real number as an eigenvalue, with eigenvector $e^{\lambda x}$.
The Characteristic Polynomial and Multiplicities
- $p_A$ 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 $E_n=\det A$, so $p_A(0)=(-1)^n\det A$.
- The eigenvalues of $A$ in $\mathbb{F}$ are exactly the roots of $p_A$ in $\mathbb{F}$; there are at most $n$ of them.
- If $p_A$ splits over $\mathbb{F}$ as $(x-\lambda_1)\cdots(x-\lambda_n)$ (always true over $\mathbb{C}$), then $\operatorname{tr}A=\lambda_1+\cdots+\lambda_n$, $\det A=\lambda_1\cdots\lambda_n$, and $E_k$ is the $k$-th elementary symmetric function of the $\lambda_i$.
- $p_{A^T}=p_A$, and $p_{P^{-1}AP}=p_A$ for invertible $P$.
- If $A=\begin{pmatrix}B&C\\0&D\end{pmatrix}$ with square blocks $B$, $D$, then $p_A=p_B\,p_D$. In particular the eigenvalues of a triangular matrix are its diagonal entries.
Proof. (1) Write the $j$-th column of $xI-A$ as $x e_j-a_j$, where $a_j$ is the $j$-th column of $A$. The determinant is linear in each column, so expanding all $n$ columns gives $2^n$ terms, one for each set $S$ of columns in which we choose $xe_j$. Such a term is $x^{|S|}$ times the determinant of the matrix whose columns are $e_j$ ($j\in S$) and $-a_j$ ($j\notin S$); expanding along the columns $e_j$, this equals $(-1)^{n-|S|}$ times the principal minor of $A$ on the rows and columns outside $S$. Grouping the terms with $|S|=n-k$ gives the coefficient $(-1)^kE_k$ of $x^{n-k}$. For $k=0$ the coefficient is $1$ (monic), for $k=1$ the $1\times1$ principal minors are the diagonal entries, and for $k=n$ we get $(-1)^n\det A$.
(2) is Theorem 9.1(e): $\lambda$ is an eigenvalue iff $p_A(\lambda)=0$; a nonzero polynomial of degree $n$ over a field has at most $n$ roots. (3) Expanding $(x-\lambda_1)\cdots(x-\lambda_n)$ gives $x^n-e_1x^{n-1}+e_2x^{n-2}-\cdots$ with $e_k$ the elementary symmetric functions; compare with (1). (4) $\det(xI-A^T)=\det\big((xI-A)^T\big)=\det(xI-A)$, and similarity was handled in Definition 9.2. (5) $xI-A=\begin{pmatrix}xI-B&-C\\0&xI-D\end{pmatrix}$ is block upper triangular, and the determinant of a block triangular matrix is the product of the determinants of the diagonal blocks (§8.3). Repeating this with $1\times1$ blocks shows $p_A(x)=\prod(x-a_{ii})$ for triangular $A$. ∎
The formula in (1) is the fastest way to get $p_A$ for a $3\times3$ matrix by hand: $p_A(x)=x^3-(\operatorname{tr}A)\,x^2+E_2\,x-\det A$, where $E_2$ is the sum of the three $2\times2$ minors that sit on the diagonal. For $2\times2$ matrices, $p_A(x)=x^2-(\operatorname{tr}A)x+\det A$.
Proof. Suppose not, and let $m$ be the smallest number of the $v_i$ that admit a nontrivial dependence; after renumbering, $c_1v_1+\cdots+c_mv_m=0$ with all $c_i\ne0$ (if some $c_i$ were $0$ we would have a shorter relation), and $m\ge2$ because eigenvectors are nonzero. Apply $T-\lambda_mI$: since $(T-\lambda_mI)v_i=(\lambda_i-\lambda_m)v_i$, we get $\sum_{i=1}^{m-1}c_i(\lambda_i-\lambda_m)v_i=0$, a relation among $m-1$ of the vectors whose coefficients $c_i(\lambda_i-\lambda_m)$ are nonzero — contradicting the minimality of $m$. 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 dependence relation, so all $w_i=0$ (§8.1, the criterion for a direct sum). ∎
Proof. $g(\lambda)\ge1$ because an eigenvalue has an eigenvector. Let $g=g(\lambda)$, choose a basis $v_1,\dots,v_g$ of $E_\lambda$ and extend it to a basis $\mathcal{B}$ of $V$ (§8.1). Since $Tv_i=\lambda v_i$ for $i\le g$, the first $g$ columns of $[T]_{\mathcal{B}}$ are $\lambda e_1,\dots,\lambda e_g$:
$$[T]_{\mathcal{B}}=\begin{pmatrix}\lambda I_g&C\\0&D\end{pmatrix}.$$By Theorem 9.2(5), $p_T(x)=(x-\lambda)^g\,p_D(x)$, so $(x-\lambda)^g$ divides $p_T$ and $g\le a(\lambda)$. The sum of the multiplicities of the roots of a polynomial of degree $n$ is at most $n$, with equality iff it is a product of linear factors. ∎
Polynomials in an Operator, Cayley–Hamilton and the Minimal Polynomial
Proof. The first two rules are immediate. For products it suffices, by bilinearity, to take $f=x^i$, $g=x^j$, where $(fg)(T)=T^{i+j}=T^iT^j$. Then $f(T)g(T)=(fg)(T)=(gf)(T)=g(T)f(T)$. If $Tv=\lambda v$, induction gives $T^kv=\lambda^kv$, and summing, $f(T)v=f(\lambda)v$. The matrix rules follow from $[ST]_{\mathcal{B}}=[S]_{\mathcal{B}}[T]_{\mathcal{B}}$ (§8.2) and $(P^{-1}AP)^k=P^{-1}A^kP$. ∎
Proof. $\dim\mathcal{L}(V)=n^2$, so the $n^2+1$ operators $I,T,\dots,T^{n^2}$ are linearly dependent: some nonzero polynomial annihilates $T$. Among the nonzero annihilating polynomials choose one of least degree and divide by its leading coefficient; its degree is $\ge1$ because $c\,I\ne0$ for $c\ne0$. If $f(T)=0$, divide: $f=qm_T+r$ with $r=0$ or $\deg r\lt\deg m_T$. Then $r(T)=f(T)-q(T)m_T(T)=0$, so by minimality $r=0$ and $m_T\mid f$. Conversely, if $f=qm_T$ then $f(T)=q(T)m_T(T)=0$. If $m$ and $m'$ are both monic of least degree, each divides the other, so they are equal. Similar matrices: $f(P^{-1}AP)=P^{-1}f(A)P$ vanishes iff $f(A)$ does. Block diagonal: $f(A)=\operatorname{diag}(f(A_1),\dots,f(A_k))$, which is $0$ iff every $f(A_i)=0$, iff every $m_{A_i}$ divides $f$, iff their least common multiple divides $f$. The statement about $p_A$ is Theorem 9.2(5) applied repeatedly. ∎
A remark that saves trouble later: $m_A$ does not change if we enlarge the field (say from $\mathbb{R}$ to $\mathbb{C}$), because the degree of $m_A$ is the first $k$ for which $A^k$ is a linear combination of $I,\dots,A^{k-1}$, and linear dependence of real matrices is decided by ranks of real matrices, which do not change over $\mathbb{C}$. Interactive 2 below finds $m_A$ in exactly this way.
Proof. Write $p_A(x)=x^n+c_{n-1}x^{n-1}+\cdots+c_1x+c_0$. Let $B(x)=\operatorname{adj}(xI-A)$, the adjugate (classical adjoint) of $xI-A$ (§8.3). Each entry of $B(x)$ is, up to sign, an $(n-1)\times(n-1)$ minor of $xI-A$, hence a polynomial in $x$ of degree at most $n-1$. Collecting powers of $x$ we can therefore write
$$B(x)=B_0+B_1x+\cdots+B_{n-1}x^{n-1}$$with constant matrices $B_k\in M_n(\mathbb{F})$. The basic property of the adjugate, $M\operatorname{adj}(M)=\det(M)\,I$, holds for matrices with entries in any commutative ring, in particular in $\mathbb{F}[x]$; with $M=xI-A$ it reads
$$(xI-A)(B_0+B_1x+\cdots+B_{n-1}x^{n-1})=(x^n+c_{n-1}x^{n-1}+\cdots+c_0)\,I.$$Two polynomials with matrix coefficients are equal iff their coefficients are equal. Comparing the coefficients of $x^n,x^{n-1},\dots,x,1$:
$$B_{n-1}=I,\qquad B_{k-1}-AB_k=c_kI\ \ (1\le k\le n-1),\qquad -AB_0=c_0I.$$Multiply the first equation on the left by $A^n$, the equation for $x^k$ by $A^k$, and the last by $I$, and add. The right-hand sides add up to $A^n+c_{n-1}A^{n-1}+\cdots+c_1A+c_0I=p_A(A)$. On the left we get
$$A^nB_{n-1}+\sum_{k=1}^{n-1}\left(A^kB_{k-1}-A^{k+1}B_k\right)-AB_0=\sum_{k=1}^{n}A^kB_{k-1}-\sum_{k=0}^{n-1}A^{k+1}B_k=0,$$because the two sums contain exactly the same terms $A^jB_{j-1}$, $j=1,\dots,n$. Hence $p_A(A)=0$. For an operator, apply this to $[T]_{\mathcal{B}}$ and use Theorem 9.5. The consequences follow from Theorem 9.6. ∎
Why not simply "substitute $x=A$ in $\det(xI-A)$ to get $\det(A-A)=0$"? Because $p_A(A)$ is a matrix, while $\det(AI-A)$ — whatever it means — is a scalar; the substitution confuses the two. The proof above is what makes the slogan correct. A second proof, via triangular forms, is given in §9.2.
Proof. If $\lambda$ is an eigenvalue with eigenvector $v$, then $0=m_T(T)v=m_T(\lambda)v$ by Theorem 9.5, and $v\ne0$ gives $m_T(\lambda)=0$. Conversely, $m_T\mid p_T$ by Cayley–Hamilton, so every root of $m_T$ is a root of $p_T$, i.e. an eigenvalue. In the split case, $m_T$ divides $\prod(x-\lambda_i)^{a_i}$, so by unique factorisation $m_T=\prod(x-\lambda_i)^{e_i}$ with $0\le e_i\le a_i$, and $e_i\ge1$ because each $\lambda_i$ is a root of $m_T$. Then $m_T^{\,n}=\prod(x-\lambda_i)^{ne_i}$ and $ne_i\ge n\ge a_i$. The general statement follows by passing to a field over which $p_T$ splits: $m_T$ is unchanged (remark after Theorem 9.6), and the roots there determine the irreducible factors over $\mathbb{F}$. ∎
- Factor $p_A$ as $\prod(x-\lambda_i)^{a_i}$. By Theorem 9.8 the candidates are $\prod(x-\lambda_i)^{e_i}$ with $1\le e_i\le a_i$ — every eigenvalue must appear.
- Test from the lowest degree up: the first candidate $f$ with $f(A)=0$ is $m_A$. If all $a_i=1$, then $m_A=p_A$ with no computation.
- Use structure. Block diagonal: $m=\operatorname{lcm}$ of the blocks (Theorem 9.6). A polynomial identity such as $A^2=nA$ gives a multiple of $m_A$ at once. $\deg m_A=1$ iff $A$ is a scalar matrix.
- Read off consequences: $A$ is invertible iff $m_A(0)\ne0$ iff $p_A(0)\ne0$; $\dim\operatorname{span}\{I,A,A^2,\dots\}=\deg m_A$.
Spectral Mapping, $AB$ versus $BA$, and Companion Matrices
- $\sigma(f(A))=f(\sigma(A))=\{f(\lambda):\lambda\in\sigma(A)\}$;
- with multiplicities, $p_{f(A)}(x)=\prod_{i=1}^n\big(x-f(\lambda_i)\big)$; hence $\operatorname{tr}f(A)=\sum f(\lambda_i)$, $\det f(A)=\prod f(\lambda_i)$, and $\operatorname{tr}(A^k)=\sum\lambda_i^k$;
- $A$ is invertible iff $0\notin\sigma(A)$; then $A^{-1}$ is a polynomial in $A$ and its eigenvalues are $\lambda_i^{-1}$.
Proof. (1) If $Av=\lambda v$, $v\ne0$, then $f(A)v=f(\lambda)v$ (Theorem 9.5), so $f(\sigma(A))\subseteq\sigma(f(A))$. Conversely let $\mu\in\sigma(f(A))$. If $f$ is constant, $f(A)=f(\lambda)I$ and there is nothing to prove. Otherwise factor over $\mathbb{C}$: $f(x)-\mu=c\prod_j(x-\beta_j)$ with $c\ne0$. Then $f(A)-\mu I=c\prod_j(A-\beta_jI)$ is not invertible, so some factor $A-\beta_jI$ is not invertible: $\beta_j\in\sigma(A)$ and $f(\beta_j)=\mu$. (2) By §9.2 (Theorem 9.17), $A=PUP^{-1}$ with $U$ upper triangular and diagonal $\lambda_1,\dots,\lambda_n$. Powers and linear combinations of upper triangular matrices are upper triangular with the corresponding powers and combinations on the diagonal, so $f(A)=Pf(U)P^{-1}$ with $f(U)$ upper triangular with diagonal $f(\lambda_1),\dots,f(\lambda_n)$; apply Theorem 9.2(3), (5). (3) $\det A=\prod\lambda_i$. If $A$ is invertible, $p_A(0)=(-1)^n\det A\ne0$, and $A^n+c_{n-1}A^{n-1}+\cdots+c_1A=-c_0I$ gives $A^{-1}=-c_0^{-1}(A^{n-1}+c_{n-1}A^{n-2}+\cdots+c_1I)$. Finally $A^{-1}=PU^{-1}P^{-1}$, and the inverse of an invertible upper triangular matrix is upper triangular with diagonal entries $\lambda_i^{-1}$, so the eigenvalues of $A^{-1}$ are the $\lambda_i^{-1}$ with the same multiplicities. ∎
Over $\mathbb{R}$ only the inclusion $f(\sigma(A))\subseteq\sigma(f(A))$ survives: the rotation $R=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ has no real eigenvalues, yet $R^2=-I$ has the real eigenvalue $-1$. The fix is to compute over $\mathbb{C}$, where $\sigma(R)=\{i,-i\}$ and $i^2=(-i)^2=-1$.
Proof. Work with matrices over the field $\mathbb{F}(x)$ of rational functions, so that $x$ is a genuine scalar. Two block multiplications give
$$\begin{pmatrix}I_m&-A\\0&I_n\end{pmatrix}\begin{pmatrix}xI_m&A\\B&I_n\end{pmatrix}=\begin{pmatrix}xI_m-AB&0\\B&I_n\end{pmatrix},\qquad\begin{pmatrix}I_m&0\\-B&xI_n\end{pmatrix}\begin{pmatrix}xI_m&A\\B&I_n\end{pmatrix}=\begin{pmatrix}xI_m&A\\0&xI_n-BA\end{pmatrix}.$$Let $\Delta$ be the determinant of the common middle factor. Taking determinants of the block triangular matrices: the first identity gives $1\cdot\Delta=p_{AB}(x)\cdot1$, the second gives $x^n\Delta=x^m\,p_{BA}(x)$. Hence $x^np_{AB}=x^mp_{BA}$. Cancelling powers of $x$ shows that the two polynomials have the same nonzero roots with the same multiplicities, and for $m=n$ that they are equal. Comparing the coefficients of $x^{n+m-1}$ gives $\operatorname{tr}(AB)=\operatorname{tr}(BA)$, and putting $x=-1$ gives $(-1)^n\det(-I_m-AB)=(-1)^m\det(-I_n-BA)$, i.e. $\det(I_m+AB)=\det(I_n+BA)$.
Geometric multiplicities: let $\lambda\ne0$. If $ABv=\lambda v$ then $BA(Bv)=\lambda Bv$, so $B$ maps $E_\lambda(AB)$ into $E_\lambda(BA)$; it is injective there, since $Bv=0$ would give $\lambda v=ABv=0$. Hence $\dim E_\lambda(AB)\le\dim E_\lambda(BA)$, and by symmetry (using $A$) equality holds. Finally, $A=\begin{pmatrix}1&0\\0&0\end{pmatrix}$, $B=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ give $AB=\begin{pmatrix}0&1\\0&0\end{pmatrix}\ne0$ and $BA=0$: $m_{AB}=x^2\ne x=m_{BA}$, and the zero matrix is similar only to itself. ∎
The most useful special case is a rank-one matrix $uv^T$ with $u,v\in\mathbb{F}^n$: here $A=u$ ($n\times1$) and $B=v^T$ ($1\times n$), and $BA=v^Tu$ is a $1\times1$ matrix. So $x\,p_{uv^T}(x)=x^n(x-v^Tu)$, i.e.
$$p_{uv^T}(x)=x^{n-1}\,(x-v^Tu),\qquad\det(I+uv^T)=1+v^Tu .$$- $p_{C(f)}=m_{C(f)}=f$. In particular, every monic polynomial of degree $n$ over $\mathbb{F}$ is the characteristic polynomial of some matrix in $M_n(\mathbb{F})$.
- For $A\in M_n(\mathbb{F})$ the following are equivalent: (a) $A$ has a cyclic vector, i.e. some $v$ with $v,Av,\dots,A^{n-1}v$ a basis of $\mathbb{F}^n$; (b) $A$ is similar to $C(p_A)$; (c) $m_A=p_A$.
- If $A$ has $n$ distinct eigenvalues in $\mathbb{F}$, then $m_A=p_A$.
Proof. (1) Write $C=C(f)$. By definition $C^ke_1=e_{k+1}$ for $0\le k\le n-1$. If $g=\sum_{k\lt n}g_kx^k\ne0$ has degree $\lt n$, then $g(C)e_1=\sum g_ke_{k+1}\ne0$; so no nonzero polynomial of degree $\lt n$ annihilates $C$, and $\deg m_C\ge n$. Next, $C^ne_1=Ce_n=-\sum_{k\lt n}c_ke_{k+1}$, so $f(C)e_1=C^ne_1+\sum_{k\lt n}c_kC^ke_1=0$. Then $f(C)e_{j}=f(C)C^{j-1}e_1=C^{j-1}f(C)e_1=0$ for every $j$ (Theorem 9.5), so $f(C)=0$ and $m_C\mid f$. Both are monic, $\deg m_C\ge n=\deg f$, hence $m_C=f$. By Cayley–Hamilton $m_C\mid p_C$, and both have degree $n$ and are monic, so $p_C=f$.
(2) (a)$\Rightarrow$(b): in the basis $\mathcal{B}=(v,Av,\dots,A^{n-1}v)$, $A$ sends each basis vector to the next, and $A^nv=-\sum_{k\lt n}d_kA^kv$ for some scalars $d_k$; so $[A]_{\mathcal{B}}=C(g)$ with $g=x^n+\sum d_kx^k$. By (1), $p_A=p_{C(g)}=g$, so $A\sim C(p_A)$. (b)$\Rightarrow$(c): similar matrices have the same $m$ and $p$, and $m=p$ for $C(p_A)$ by (1). (c)$\Rightarrow$(a), when $\mathbb{F}$ is infinite: for each $v$ the polynomials $f$ with $f(A)v=0$ are the multiples of a monic $m_v$ (same argument as Theorem 9.6), and $m_v\mid m_A$. The monic divisors $d$ of $m_A$ are finitely many, and $V=\bigcup_d\ker d(A)$ because $v\in\ker m_v(A)$. A vector space over an infinite field is not a finite union of proper subspaces (§8.1), so $\ker d(A)=V$ for some divisor $d$; then $d(A)=0$, so $m_A\mid d$ and $d=m_A$. Thus some $v$ has $m_v=m_A=p_A$ of degree $n$, which means that $v,Av,\dots,A^{n-1}v$ are independent. (For finite fields the statement remains true; it follows from the rational canonical form, §9.3.) (3) All $n$ eigenvalues are roots of $m_A$ (Theorem 9.8), so $\deg m_A\ge n$. ∎
- Eigenvalues live in the field: $\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ has no real eigenvalue but the eigenvalues $\pm i$ over $\mathbb{C}$. A real matrix of odd size always has a real eigenvalue.
- Eigenvectors are nonzero, eigenvalues may be $0$: $0\in\sigma(A)\iff A$ is singular.
- $g(\lambda)$ can be smaller than $a(\lambda)$: $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ has $a(1)=2$, $g(1)=1$.
- Same $p_A$ does not mean similar: $I_2$ and $\begin{pmatrix}1&1\\0&1\end{pmatrix}$. Same $p_A$ and $m_A$ is enough for $n\le3$ but not for $n=4$: the $4\times4$ matrices with two nilpotent $2\times2$ blocks, and with one such block and two zeros, both have $p=x^4$, $m=x^2$, but ranks $2$ and $1$.
- $m_A=p_A$ does not mean distinct eigenvalues: $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ has $m=p=(x-1)^2$. Only the converse (distinct $\Rightarrow m=p$) holds.
- The naive proof of Cayley–Hamilton is wrong: "$p_A(A)=\det(AI-A)=0$" equates a matrix with a scalar.
- $A$ and $A^T$: same eigenvalues and multiplicities, different eigenvectors ($\begin{pmatrix}1&1\\0&0\end{pmatrix}$).
- Eigenvalues are not additive or multiplicative: $N_1=\begin{pmatrix}0&1\\0&0\end{pmatrix}$, $N_2=N_1^T$ have only the eigenvalue $0$, but $N_1+N_2$ has eigenvalues $\pm1$ and $N_1N_2=\operatorname{diag}(1,0)$ has eigenvalue $1$.
- Row operations change eigenvalues: $\begin{pmatrix}1&2\\3&4\end{pmatrix}$ has eigenvalues $\tfrac{5\pm\sqrt{33}}2$; its row-echelon form $\begin{pmatrix}1&2\\0&-2\end{pmatrix}$ has eigenvalues $1,-2$. Only $\det$ and rank survive row reduction.
- Rank is not the number of nonzero eigenvalues: $\begin{pmatrix}0&1\\0&0\end{pmatrix}$ has rank $1$ and no nonzero eigenvalue. In general $\operatorname{rank}A\ge$ (number of nonzero eigenvalues counted with algebraic multiplicity), with equality for diagonalisable $A$ (§9.2).
- Spectral mapping over $\mathbb{R}$: $R^2=-I$ has eigenvalue $-1$ though the rotation $R$ has no real eigenvalue. Work over $\mathbb{C}$.
- $AB$ versus $BA$: same characteristic polynomial (square case) but possibly different minimal polynomials; for $A$ of size $m\times n$ with $m\gt n$, $AB$ has at least $m-n$ extra zero eigenvalues.
- Infinite dimension: the right shift has no eigenvalues although it is not invertible; multiplication by $t$ on $C[0,1]$ has no eigenvalues; $D$ on $C^\infty(\mathbb{R})$ has every real eigenvalue.
Exercises
- 1 Find the eigenvalues and eigenvectors of $\begin{pmatrix}4&-2\\1&1\end{pmatrix}$.
- 2 Find the eigenvalues of $\begin{pmatrix}0&1\\-1&0\end{pmatrix}$ over $\mathbb{R}$ and over $\mathbb{C}$, with eigenvectors in the complex case.
- 3 Find the eigenvalues and an eigenvector for each of $\begin{pmatrix}1&2&3\\0&4&5\\0&0&6\end{pmatrix}$.
- 4 Find $a(\lambda)$ and $g(\lambda)$ for every eigenvalue of $\begin{pmatrix}2&1&0\\0&2&0\\0&0&2\end{pmatrix}$.
- 5 Prove: if $A^2=A$ every eigenvalue is $0$ or $1$; if $A^k=0$ every eigenvalue is $0$; if $A^2=I$ every eigenvalue is $\pm1$.
- 6 Let $A$ be invertible and $c\in\mathbb{F}$. Show that $v$ is an eigenvector of $A$ for $\lambda$ iff it is an eigenvector of $A^{-1}$ for $\lambda^{-1}$, iff it is an eigenvector of $A+cI$ for $\lambda+c$.
- 7 Find the eigenvalues, with multiplicities, and the determinant of the $n\times n$ matrix with every diagonal entry $a$ and every off-diagonal entry $b$.
- 8 Find the eigenvalues and eigenvectors of $T(p)=x\,p'(x)$ on $P_n$.
- 1 Use principal minors to find $p_A$ for $A=\begin{pmatrix}2&1&1\\1&2&1\\1&1&2\end{pmatrix}$, and factor it.
- 2 $A\in M_3(\mathbb{R})$ has eigenvalues $1,-1,2$. Find $\operatorname{tr}A$, $\det A$, $\operatorname{tr}(A^2)$, $\det(A^2+A)$ and the eigenvalues of $\operatorname{adj}A$.
- 3 Show that a $2\times2$ matrix $A$ has a repeated eigenvalue iff $(\operatorname{tr}A)^2=4\det A$.
- 4 A real $2\times2$ matrix has trace $2$ and determinant $5$. Find its eigenvalues.
- 5 Let $n$ be even and $A\in M_n(\mathbb{R})$ with $\det A\lt0$. Prove that $A$ has at least one positive and at least one negative eigenvalue.
- 6 Prove that a nilpotent $A\in M_n(\mathbb{F})$ has $p_A=x^n$, and deduce $\operatorname{tr}A=0$ and $\det(I+A)=1$.
- 7 Let $A\in M_3(\mathbb{C})$ with $\operatorname{tr}A=\operatorname{tr}A^2=\operatorname{tr}A^3=0$. Prove that $A$ is nilpotent.
- 8 Prove that $\operatorname{tr}(AB)=\operatorname{tr}(BA)$ directly from the definition, and deduce that there are no $A,B\in M_n(\mathbb{R})$ with $AB-BA=I$.
- 1 Verify the Cayley–Hamilton theorem for $\begin{pmatrix}1&2\\3&4\end{pmatrix}$.
- 2 Use Cayley–Hamilton to find $M^{-1}$ for $M=\begin{pmatrix}2&0&1\\1&1&0\\0&1&1\end{pmatrix}$.
- 3 Compute $A^{10}$ for $A=\begin{pmatrix}1&1\\0&2\end{pmatrix}$ without multiplying ten matrices.
- 4 Find the minimal polynomials of $\begin{pmatrix}2&1&0&0\\0&2&0&0\\0&0&2&1\\0&0&0&2\end{pmatrix}$ and of $\operatorname{diag}(1,2,2,3)$.
- 5 Show that a projection $E\ne0,I$ has $m_E=x(x-1)$, and a nilpotent matrix of index $k$ has $m=x^k$.
- 6 Prove that $A$ is invertible iff $m_A(0)\ne0$, and that then $A^{-1}$ is a polynomial in $A$ of degree $\deg m_A-1$.
- 7 Prove that $\dim\operatorname{span}\{I,A,A^2,\dots\}=\deg m_A$. Deduce that this span has dimension at most $n$.
- 8 Show that there is no $A\in M_3(\mathbb{R})$ with $A^2=-I$, but that there is one in $M_2(\mathbb{R})$. What are the possible minimal polynomials of a real $A$ with $A^2=-I$?
- 1 $A\in M_3(\mathbb{C})$ has eigenvalues $1,2,-1$. Find the eigenvalues, trace and determinant of $A^3-2A+I$.
- 2 Find the eigenvalues of $uv^T$ and of $I+uv^T$ for $u=(1,2,3)^T$, $v=(1,1,1)^T$, and $\det(I+uv^T)$.
- 3 Let $A$ be $5\times3$ and $B$ be $3\times5$, with $BA$ having eigenvalues $1,2,3$. Find $p_{AB}$, $\operatorname{tr}(AB)$ and $\det(I_5+AB)$.
- 4 Write down the companion matrix of $x^4-1$ and find its eigenvalues. Which familiar matrix is it?
- 5 Let $P$ be the $n\times n$ cyclic permutation matrix $Pe_j=e_{j+1}$ (indices mod $n$) and $Q=c_0I+c_1P+\cdots+c_{n-1}P^{n-1}$ (a circulant). Show that the eigenvalues of $Q$ are $f(\omega^k)$, $k=0,\dots,n-1$, where $f(x)=\sum c_jx^j$ and $\omega=e^{2\pi i/n}$. Check with $\begin{pmatrix}2&1&1\\1&2&1\\1&1&2\end{pmatrix}$.
- 6 Prove: if $A$ has a cyclic vector, every matrix commuting with $A$ is a polynomial in $A$.
- 7 Prove or disprove: if $A\in M_n(\mathbb{C})$ and $A^2$ has $n$ distinct eigenvalues, then so does $A$.
- 8 Applied. A fish population has juveniles and adults. Each year, each juvenile produces $1$ offspring and each adult $1.5$ offspring, half of the juveniles survive to become adults, and no adult survives. With $L=\begin{pmatrix}1&1.5\\0.5&0\end{pmatrix}$ and $x_{k+1}=Lx_k$, find the long-run annual growth factor and the long-run proportion of juveniles.
E1. CSIR Part B GATE MA Let $A\in M_3(\mathbb{R})$ satisfy $\operatorname{tr}A=6$, $\operatorname{tr}(A^2)=14$ and $\det A=6$. Then $\det(A+I)$ equals (A) $6$; (B) $12$; (C) $24$; (D) $30$.
E2. CSIR Part B Let $A$ be a real $3\times2$ matrix and $B$ a real $2\times3$ matrix with $BA=\begin{pmatrix}1&2\\2&1\end{pmatrix}$. The characteristic polynomial of $AB$ is (A) $(x-3)(x+1)$; (B) $x(x-3)(x+1)$; (C) $x^2(x-3)$; (D) not determined by the given data.
E3. CSIR Part B Let $M=J_5-I_5$, where $J_5$ is the $5\times5$ matrix with all entries $1$. The minimal polynomial of $M$ is (A) $(x-4)(x+1)$; (B) $(x-4)(x+1)^4$; (C) $x(x-5)$; (D) $(x-5)(x+1)$.
E4. CSIR Part C Let $A\in M_3(\mathbb{R})$ have characteristic polynomial $x^3-x$. Which of the following are true? (One or more options may be correct.) (A) $A$ is invertible; (B) $A^2+I$ is invertible; (C) $A^5=A$; (D) the minimal polynomial of $A$ is $x^3-x$.
E5. CSIR Part C GATE MA Let $A,B\in M_n(\mathbb{R})$, $n\ge2$. Which of the following are necessarily true? (A) $AB$ and $BA$ have the same eigenvalues; (B) $AB$ and $BA$ have the same minimal polynomial; (C) $\operatorname{tr}\big((AB)^k\big)=\operatorname{tr}\big((BA)^k\big)$ for every $k\ge1$; (D) $AB$ and $BA$ are similar.
E6. CSIR Part C Let $T:M_2(\mathbb{R})\to M_2(\mathbb{R})$ be $T(X)=X+X^{T}$. Which of the following are true? (A) $0$ is an eigenvalue of $T$ with geometric multiplicity $1$; (B) $\det T=8$; (C) $T^2=2T$; (D) the minimal polynomial of $T$ is $x(x-2)^3$.
Suggested plan — 2 lectures × 60 min (a third for practice).
- Eigenvalues and the characteristic polynomial. Definition 9.1 and Theorem 9.1, with the infinite-dimensional counterexamples kept to two minutes. Open Interactive 1 with $[3\ 1;\ 0\ 2]$: the zeros of $g(\theta)$ are the eigen-directions; then the shear (double zero), the rotation (no zero) and the scalar matrix ($g\equiv0$). Theorem 9.2 with the principal-minor formula and Example 3 on the board; Theorems 9.3 and 9.4 in full; Example 4 as a quick poll.
- Cayley–Hamilton and the minimal polynomial. Polynomials in $T$, Theorem 9.6, then the adjugate proof of Cayley–Hamilton (about 15 minutes — write the coefficient comparison as a column so that the telescoping is visible). Theorem 9.8 and the strategy box; Examples 5 and 6. Project Interactive 2 — first the two presets $\operatorname{diag}(2,2,3)$ and $[2\ 1\ 0;\dots]$, then the two $4\times4$ nilpotents. Close with Theorems 9.10 and 9.11 (statements and the block-matrix trick) and E2 or E5.
Discussion questions. (1) Over which fields does every $2\times2$ matrix have an eigenvalue? Every $3\times3$ matrix? (2) Why does the proof of Cayley–Hamilton need the adjugate rather than the inverse of $xI-A$? (3) The minimal polynomial is unchanged by passing from $\mathbb{R}$ to $\mathbb{C}$; is the same true of the eigenvalues? Of the eigenvectors?
Misconceptions to watch for. Computing eigenvalues after row reduction; forgetting that every eigenvalue must appear in $m_A$; believing $m_A=p_A$ forces distinct eigenvalues; treating $\det(A-A)=0$ as a proof of Cayley–Hamilton; assuming $\sigma(A+B)=\sigma(A)+\sigma(B)$; confusing "rank" with "number of nonzero eigenvalues"; dropping the extra zero eigenvalues of $AB$ when $A$ is not square.
Suggested homework. I(1, 3, 5, 7), II(1, 2, 5, 6, 7), III(2, 3, 4, 6, 8), IV(1, 3, 5, 6, 8); Exam corner E1, E4, E6.
Research question. Is a harvested fish stock growing or shrinking — and which single life-history number would a fisheries manager most like to change?
Why it matters. Age-structured (Leslie) models are used by fisheries agencies in the North Atlantic and the Indian Ocean, by conservation biologists for sea turtles and albatrosses, and by demographers for human populations. The long-run growth factor is the dominant eigenvalue of the Leslie matrix, the stable age distribution is its eigenvector, and the "sensitivities" of the eigenvalue tell managers whether protecting eggs, juveniles or breeding adults matters most.
What to do.
- For three age classes take fertilities $f=(0,\,1.5,\,1.2)$ and survival rates $s_1=0.5$, $s_2=0.4$, so $L=\begin{pmatrix}0&1.5&1.2\\0.5&0&0\\0&0.4&0\end{pmatrix}$. Show that $p_L(x)=x^3-f_1x^2-f_2s_1x-f_3s_1s_2$ — compare with the companion matrix of Definition 9.5.
- Find the dominant root $\lambda_1$ numerically. Is the stock growing?
- Check your answer by iterating $x_{k+1}=Lx_k$ from $x_0=(100,0,0)$ for $40$ years and computing $\|x_{k+1}\|/\|x_k\|$.
- Find the value of $f_2$ that makes $\lambda_1=1$ exactly (a stationary stock), and compare it with the effect of raising $s_1$ instead.
Tools: Python (numpy.roots, numpy.linalg.eig) or a spreadsheet.
Hand in: A one-page note with $p_L$, $\lambda_1$, the stable age distribution, your iteration plot, and a recommendation in one sentence.
Going further. Read about the Perron–Frobenius theorem, which guarantees a positive dominant eigenvalue with a positive eigenvector for such matrices, and about elasticity analysis in population biology.
🧪 Self-check quiz
Q1. The eigenvalues of $\begin{pmatrix}2&5\\0&3\end{pmatrix}$ are
Q2. $A\in M_3(\mathbb{R})$ has eigenvalues $1,2,4$. Then $\det(A^{-1})$ is
Q3. If $p_A(x)=x^2-5x+6$, then $A^{-1}$ equals
Q4. The minimal polynomial of the $3\times3$ matrix with all entries $1$ is
Q5. Which statement holds for every eigenvalue $\lambda$ of every $A\in M_n(\mathbb{C})$?