Eigenvalues and the Cayley–Hamilton Theorem
CSIR NET linear algebra: characteristic and minimal polynomials, trace and determinant, eigenvalues of special matrices, AB vs BA, Gershgorin, Cayley–Hamilton.
- 18 min
On this page · 11 sections
Prerequisites
- Rank, Determinants and Linear Equations: \( \lambda \) is an eigenvalue iff \( A - \lambda I \) is singular, i.e. \( \det(A - \lambda I) = 0 \).
- Vector Spaces, Bases and Linear Transformations: eigenspaces are kernels, so their dimensions come from rank–nullity.
- Roots of polynomials over \( \mathbb{C} \) and the fact that real polynomials have conjugate-pair roots.
The Hook: Fibonacci in a Matrix
Eigenvalues and the Characteristic Polynomial
Writing \( \chi_A(x) = x^n - c_1x^{n-1} + c_2x^{n-2} - \dots + (-1)^nc_n \):
- \( c_1 = \operatorname{tr} A = \sum \lambda_i \);
- \( c_n = \det A = \prod \lambda_i \);
- \( c_k \) = sum of all \( k \times k \) principal minors = \( k \)-th elementary symmetric function of the eigenvalues.
For a \( 2 \times 2 \) matrix, \( \chi_A(x) = x^2 - (\operatorname{tr} A)x + \det A \). For a \( 3 \times 3 \), \( \chi_A(x) = x^3 - (\operatorname{tr} A)x^2 + (\text{sum of the three } 2 \times 2 \text{ principal minors})x - \det A \).
Multiplicities
- Algebraic multiplicity \( a(\lambda) \): the multiplicity of \( \lambda \) as a root of \( \chi_A \).
- Geometric multiplicity \( g(\lambda) = \dim E_\lambda = n - \operatorname{rank}(A - \lambda I) \).
- Always \( 1 \le g(\lambda) \le a(\lambda) \). Equality for every eigenvalue (with \( \chi_A \) splitting) is exactly diagonalisability; see Change of Basis and Canonical Forms.
- Eigenvectors for distinct eigenvalues are linearly independent. Hence \( n \) distinct eigenvalues force diagonalisability.
Eigenvalues of related matrices
If \( A \) has eigenvalues \( \lambda_1, \dots, \lambda_n \) (over \( \mathbb{C} \), with multiplicity):
| Matrix | Eigenvalues | Note |
|---|---|---|
| \( A^T \) | same \( \lambda_i \) | eigenvectors differ in general |
| \( \bar{A} \), \( A^* \) | \( \bar{\lambda}_i \) | |
| \( A^k \) | \( \lambda_i^k \) | |
| \( p(A) \), \( p \) a polynomial | \( p(\lambda_i) \) | spectral mapping |
| \( A^{-1} \) | \( \frac{1}{\lambda_i} \) | same eigenvectors |
| \( A + cI \) | \( \lambda_i + c \) | |
| \( \operatorname{adj} A \) (\( A \) invertible) | \( \frac{\det A}{\lambda_i} \) | \( \operatorname{adj} A = (\det A)A^{-1} \) |
| \( P^{-1}AP \) | same \( \lambda_i \) | similarity invariant |
| \( BA \) vs \( AB \) | same non-zero eigenvalues | see below |
So \( \det p(A) = \prod p(\lambda_i) \) and \( \operatorname{tr} p(A) = \sum p(\lambda_i) \): no need to form \( p(A) \).
\( AB \) and \( BA \) need not be similar: \( A = E_{11} \), \( B = E_{12} \) give \( AB = E_{12} \ne 0 \) and \( BA = 0 \).
Eigenvalues of special matrices
| Matrix type | Eigenvalues lie in | Extra |
|---|---|---|
| triangular | the diagonal entries | |
| real symmetric / Hermitian | \( \mathbb{R} \) | orthonormal eigenbasis |
| real skew-symmetric / skew-Hermitian | \( i\mathbb{R} \) | real skew-symmetric of odd order is singular |
| real orthogonal / unitary | the unit circle | \( \det = \pm 1 \) (orthogonal) |
| nilpotent (\( N^k = 0 \)) | \( \{ 0 \} \) | \( \operatorname{tr} N = 0 \), \( \chi_N = x^n \) |
| idempotent (\( P^2 = P \)) | \( \{ 0, 1 \} \) | \( \operatorname{tr} P = \operatorname{rank} P \) |
| involution (\( A^2 = I \)) | \( \{ 1, -1 \} \) | diagonalisable if \( \operatorname{char} F \ne 2 \) |
| \( A^k = I \) | \( k \)-th roots of unity | diagonalisable over \( \mathbb{C} \) |
| rank one, \( uv^T \) | \( v^Tu \) and \( 0 \) (\( n - 1 \) times) | |
| \( J_n \) (all ones) | \( n \) and \( 0 \) (\( n - 1 \) times) | |
| \( aI + bJ_n \) | \( a + nb \) and \( a \) (\( n - 1 \) times) | |
| stochastic (rows sum to 1) | 1 is an eigenvalue | eigenvector \( (1, \dots, 1)^T \) |
General principle: if \( p(A) = 0 \) for a polynomial \( p \), every eigenvalue \( \lambda \) satisfies \( p(\lambda) = 0 \) (because \( p(A)v = p(\lambda)v \)). The converse fails: the eigenvalues of \( p(A) \) being zero does not make \( p(A) = 0 \).
Real matrices
- Non-real eigenvalues of a real matrix come in conjugate pairs.
- A real matrix of odd order has at least one real eigenvalue (\( \chi_A \) is a real polynomial of odd degree).
- A real \( n \times n \) matrix may have no real eigenvalue when \( n \) is even: the rotation \( \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} \) has eigenvalues \( \pm i \).
Gershgorin’s theorem
Consequences: a strictly diagonally dominant matrix (\( \lvert a_{ii} \rvert > R_i \) for all \( i \)) is invertible, because 0 lies in no disc. For \( \begin{pmatrix} 10 & 1 & 1 \\ 0.5 & 5 & 0.5 \\ 1 & 1 & -3 \end{pmatrix} \) the discs are centred at \( 10, 5, -3 \) with radii \( 2, 1, 2 \): pairwise disjoint, so each holds exactly one eigenvalue. For a real matrix a disc holding a single eigenvalue must hold a real one (the conjugate would sit in the same disc), so all three eigenvalues are real, one in each of \( [8, 12] \), \( [4, 6] \), \( [-5, -1] \).
The Cayley–Hamilton Theorem and the Minimal Polynomial
Uses.
- Inverse. If \( \chi_A(x) = x^n + a_{n-1}x^{n-1} + \dots + a_1x + a_0 \) with \( a_0 = (-1)^n\det A \ne 0 \), then $$A^{-1} = -\frac{1}{a_0}\left(A^{n-1} + a_{n-1}A^{n-2} + \dots + a_1I\right).$$ For \( A = \begin{pmatrix} 1 & 2 \\ 3 & 4 \end{pmatrix} \): \( A^2 - 5A - 2I = 0 \), so \( A(A - 5I) = 2I \) and \( A^{-1} = \frac{1}{2}A - \frac{5}{2}I \).
- Powers. Divide \( x^N \) by \( \chi_A(x) \): \( x^N = q(x)\chi_A(x) + r(x) \) with \( \deg r < n \), so \( A^N = r(A) \). To find \( r \) for a \( 2 \times 2 \) matrix with distinct eigenvalues \( \lambda, \mu \), solve \( r(\lambda) = \lambda^N \), \( r(\mu) = \mu^N \).
- Every power of \( A \) is a polynomial in \( A \) of degree \( < n \). So the matrices \( I, A, A^2, \dots \) span a space of dimension at most \( n \).
| Property of \( m_A \) | Consequence |
|---|---|
| \( m_A \mid \chi_A \) | \( \deg m_A \le n \) |
| same roots as \( \chi_A \) (multiplicities may drop) | every eigenvalue is a root of \( m_A \) |
| multiplicity of \( \lambda \) in \( m_A \) = size of the largest Jordan block for \( \lambda \) | read off the Jordan form |
| \( A \) diagonalisable \( \iff m_A \) splits with distinct roots | e.g. \( A^2 = A \) or \( A^3 = A \) \( \Rightarrow \) diagonalisable |
| similar matrices have the same \( m_A \) | |
| \( \deg m_A = n \iff \chi_A = m_A \) | then \( A \) is similar to the companion matrix of \( \chi_A \) |
Finding \( m_A \). List the divisors of \( \chi_A \) that contain every eigenvalue, try them from lowest degree up. For \( A = \begin{pmatrix} 2 & 1 & 0 \\ 0 & 2 & 0 \\ 0 & 0 & 2 \end{pmatrix} \): \( \chi_A = (x - 2)^3 \); \( A - 2I = E_{12} \ne 0 \) but \( (A - 2I)^2 = 0 \), so \( m_A = (x - 2)^2 \).
Companion matrix. The companion matrix of \( x^n + a_{n-1}x^{n-1} + \dots + a_0 \) has that polynomial as both its characteristic and minimal polynomial. So every monic polynomial of degree \( n \) is the characteristic polynomial of some \( n \times n \) matrix, a fact Part C uses for existence questions (“there is a real \( 3 \times 3 \) matrix with \( \chi_A = x^3 - 2 \)”: true).
graph TD
A["Given a relation p(A) = 0"] --> B["Every eigenvalue is a root of p"]
A --> C["m_A divides p"]
C --> D{"p has distinct roots
in the field?"}
D -->|yes| E["A is diagonalisable over that field"]
D -->|no| F["Need more information
(e.g. A^2 = 0 allows non-diagonalisable A)"]
B --> G["tr A, det A from the possible
eigenvalue multisets"]Memory Tricks
- Trace adds, determinant multiplies. \( \operatorname{tr} = \sum \lambda \), \( \det = \prod \lambda \), and \( p(A) \) just applies \( p \) to each \( \lambda \).
- AB and BA share the non-zero part. Same non-zero eigenvalues, same trace; the larger one pads with zeros.
- Gershgorin: centre on the diagonal, radius from the row. Disjoint discs isolate eigenvalues.
- Distinct roots unlock diagonalisation. If \( A \) is killed by a polynomial with distinct roots, \( A \) is diagonalisable.
- g between 1 and a. Geometric multiplicity \( 1 \le g \le a \), algebraic multiplicity on the outside.
Don’t Confuse
| This | Is not this | The difference |
|---|---|---|
| eigenvalues of \( A^T \) = those of \( A \) | eigenvectors of \( A^T \) = those of \( A \) | eigenvectors of \( A^T \) are left eigenvectors of \( A \) |
| \( AB \), \( BA \) have the same eigenvalues | \( AB \), \( BA \) are similar | \( E_{11}E_{12} \ne 0 = E_{12}E_{11} \) |
| \( m_A \) | \( \chi_A \) | \( m_A \) divides \( \chi_A \) and has the same roots, possibly lower multiplicities |
| all eigenvalues 0 | \( A = 0 \) | nilpotent \( E_{12} \ne 0 \) |
| all eigenvalues 1 | \( A = I \) | \( \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \); equality needs diagonalisability |
| algebraic multiplicity | geometric multiplicity | root multiplicity vs \( \dim \ker(A - \lambda I) \) |
| \( \det(A + B) \) | \( \det A + \det B \) | eigenvalues of \( A + B \) are not sums of eigenvalues unless \( A, B \) are simultaneously triangularisable |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C, one or more options may be correct.
Level 1: Foundation
- Eigenvalues of \( \begin{pmatrix} 2 & 1 \\ 0 & 3 \end{pmatrix} \)?
- \( A \) has eigenvalues \( 1, 2, 3 \). Find \( \det A \), \( \operatorname{tr} A \) and the eigenvalues of \( A^{-1} \).
- State the Cayley–Hamilton theorem.
Answers
- 2 and 3 (triangular: read the diagonal).
- \( \det A = 6 \), \( \operatorname{tr} A = 6 \); eigenvalues of \( A^{-1} \): \( 1, \frac{1}{2}, \frac{1}{3} \).
- Every square matrix \( A \) over a field satisfies \( \chi_A(A) = 0 \), where \( \chi_A(x) = \det(xI - A) \).
Level 2: Part B Style
Practice Q1. A real \( 3 \times 3 \) matrix \( A \) has eigenvalues \( 1, -1, 2 \). Then \( \det(A^2 + I) \) is:
(a) 4 (b) 6 (c) 10 (d) 20
Answer: (d)
\( A^2 + I \) has eigenvalues \( \lambda^2 + 1 \): \( 2, 2, 5 \). Product \( 20 \).
Practice Q2. Let \( J \) be the \( 4 \times 4 \) matrix with every entry 1. Then \( \operatorname{tr}(J^5) \) is:
(a) 4 (b) 20 (c) 256 (d) 1024
Answer: (d)
\( J \) has eigenvalues \( 4, 0, 0, 0 \), so \( J^5 \) has \( 4^5, 0, 0, 0 \) and trace \( 1024 \). (Equivalently \( J^2 = 4J \), so \( J^5 = 4^4J \), trace \( 256 \cdot 4 \).)
Practice Q3. For \( A = \begin{pmatrix} 1 & 2 \\ 3 & 4 \end{pmatrix} \), \( A^{-1} = \alpha A + \beta I \) where \( (\alpha, \beta) \) is:
(a) \( \left(\frac{1}{2}, -\frac{5}{2}\right) \) (b) \( \left(-\frac{1}{2}, \frac{5}{2}\right) \) (c) \( (1, -5) \) (d) \( \left(\frac{1}{2}, \frac{5}{2}\right) \)
Answer: (a)
\( \operatorname{tr} A = 5 \), \( \det A = -2 \), so \( A^2 - 5A - 2I = 0 \). Then \( A(A - 5I) = 2I \) and \( A^{-1} = \frac{1}{2}A - \frac{5}{2}I \). Check: \( \frac{1}{2}A - \frac{5}{2}I = \begin{pmatrix} -2 & 1 \\ 1.5 & -0.5 \end{pmatrix} \), and \( A \) times this is \( I \).
Practice Q4. The minimal polynomial of \( \begin{pmatrix} 2 & 1 & 0 \\ 0 & 2 & 0 \\ 0 & 0 & 2 \end{pmatrix} \) is:
(a) \( x - 2 \) (b) \( (x - 2)^2 \) (c) \( (x - 2)^3 \) (d) \( x^2 - 4 \)
Answer: (b)
\( A - 2I = E_{12} \ne 0 \) and \( E_{12}^2 = 0 \). The largest Jordan block has size 2.
Level 3: Part C Style
Practice Q5. Let \( A \in M_n(\mathbb{C}) \) with \( A^2 = A \). Which statements are true?
(a) Every eigenvalue of \( A \) is 0 or 1. (b) \( A \) is diagonalisable. (c) \( \operatorname{tr} A = \operatorname{rank} A \). (d) \( I + A \) is invertible.
Answer: (a), (b), (c) and (d)
- (a) True. \( Av = \lambda v \Rightarrow \lambda v = A^2v = \lambda^2v \), so \( \lambda^2 = \lambda \).
- (b) True. \( m_A \) divides \( x(x - 1) \), which has distinct roots.
- (c) True. Diagonalised, \( A \) is \( \operatorname{diag}(1, \dots, 1, 0, \dots, 0) \); the number of 1s is both the trace and the rank.
- (d) True. The eigenvalues of \( I + A \) are 1 or 2, never 0. Explicitly \( (I + A)\left(I - \frac{1}{2}A\right) = I + \frac{1}{2}A - \frac{1}{2}A^2 = I \).
Practice Q6. Let \( A \) be a real \( 2 \times 3 \) matrix and \( B \) a real \( 3 \times 2 \) matrix such that \( AB \) has eigenvalues 1 and 2. Which statements are true?
(a) \( \det(BA) = 0 \). (b) \( \operatorname{tr}(BA) = 3 \). (c) \( BA \) is diagonalisable. (d) \( \operatorname{rank}(BA) = 3 \).
Answer: (a), (b) and (c)
\( \chi_{BA}(x) = x \, \chi_{AB}(x) = x(x - 1)(x - 2) \).
- (a) True. 0 is an eigenvalue of \( BA \). (Also \( \operatorname{rank}(BA) \le 2 < 3 \).)
- (b) True. \( 0 + 1 + 2 = 3 \).
- (c) True. Three distinct eigenvalues.
- (d) False. \( \operatorname{rank}(BA) \le \operatorname{rank} A \le 2 \).
Practice Q7. Let \( A \) be a real \( 3 \times 3 \) matrix with \( A^3 = I \) and \( A \ne I \). Which statements are true?
(a) 1 is an eigenvalue of \( A \). (b) \( A \) is diagonalisable over \( \mathbb{C} \). (c) \( \operatorname{tr} A = 1 \). (d) \( A \) is not diagonalisable over \( \mathbb{R} \).
Answer: (a), (b) and (d)
The eigenvalues are cube roots of unity: \( 1, \omega, \omega^2 \) with \( \omega = e^{2\pi i/3} \). Since \( m_A \mid x^3 - 1 \), which has distinct roots, \( A \) is diagonalisable over \( \mathbb{C} \): (b) true. If all three eigenvalues were 1, diagonalisability would give \( A = I \), excluded. A real matrix has non-real eigenvalues in conjugate pairs, and an odd-order real matrix has a real eigenvalue, which must be 1. So the eigenvalues are exactly \( 1, \omega, \omega^2 \).
- (a) True. As shown.
- (c) False. \( \operatorname{tr} A = 1 + \omega + \omega^2 = 0 \).
- (d) True. \( \omega \notin \mathbb{R} \), so there is no real eigenbasis. (An example: the permutation matrix of the 3-cycle.)
Quick Revision Table
| Idea | Fact |
|---|---|
| \( \chi_A \) coefficients | \( \operatorname{tr} A \), sums of principal minors, \( \det A \) |
| Spectral mapping | eigenvalues of \( p(A) \) are \( p(\lambda_i) \) |
| \( A^{-1} \), \( \operatorname{adj} A \) | \( \frac{1}{\lambda_i} \), \( \frac{\det A}{\lambda_i} \) |
| \( AB \) vs \( BA \) | \( \chi_{BA} = x^{n-m}\chi_{AB} \); not necessarily similar |
| Multiplicities | \( 1 \le g(\lambda) \le a(\lambda) \); \( g = n - \operatorname{rank}(A - \lambda I) \) |
| Special matrices | symmetric real; skew imaginary; orthogonal on unit circle; nilpotent 0; idempotent 0, 1 |
| Rank one \( uv^T \) | eigenvalues \( v^Tu \), 0, …, 0 |
| Odd-order real matrix | has a real eigenvalue |
| Gershgorin | eigenvalues in discs centred at \( a_{ii} \), radius = off-diagonal row sum |
| Cayley–Hamilton | \( \chi_A(A) = 0 \); gives \( A^{-1} \) and high powers |
| Minimal polynomial | divides every annihilating polynomial, same roots as \( \chi_A \) |
| Diagonalisable | \( \iff m_A \) has distinct roots (in the field) |
Key Takeaways
What’s Next
- Change of Basis and Canonical Forms: diagonal, triangular and Jordan forms, built from eigenvalues and the minimal polynomial.
- Rank, Determinants and Linear Equations to revise the determinant facts used here.
- Inner Product Spaces and Orthonormal Bases: why symmetric matrices have real eigenvalues and orthogonal eigenvectors.
- Linear Algebra Quick Revision.