Change of Basis and Canonical Forms
CSIR NET linear algebra: change of basis, similarity, diagonalisability tests, triangular (Schur) form, Jordan form from ranks, counting Jordan forms.
- 19 min
On this page · 13 sections
Prerequisites
- Eigenvalues and the Cayley–Hamilton Theorem: characteristic and minimal polynomials, algebraic and geometric multiplicity.
- Vector Spaces, Bases and Linear Transformations: the matrix of a linear map depends on the basis.
- Rank, Determinants and Linear Equations: the Jordan form is read off from ranks of powers.
The Hook: Same Map, Different Pictures
Change of Basis
Let \( B = (v_1, \dots, v_n) \) and \( B' = (v_1', \dots, v_n') \) be bases of \( V \). The change-of-basis matrix \( P \) has as its \( j \)-th column the \( B \)-coordinates of \( v_j' \). Then:
$$[v]_B = P\,[v]_{B'}, \qquad \boxed{[T]_{B'} = P^{-1}\,[T]_B\,P}.$$Coordinates transform with \( P^{-1} \) (\( [v]_{B'} = P^{-1}[v]_B \)), while the basis vectors themselves are given by \( P \): the “contravariant” behaviour that makes students mix up \( P \) and \( P^{-1} \). The safest habit: put the new basis vectors as columns of \( P \), then use \( P^{-1}AP \).
Similarity. \( A \sim B \) if \( B = P^{-1}AP \) for some invertible \( P \): two matrices are similar iff they represent the same linear map in two bases.
| Similarity invariant | Not a similarity invariant |
|---|---|
| rank, nullity | the entries, the eigenvectors |
| trace, determinant | symmetry (\( A = A^T \)) |
| characteristic and minimal polynomial | being diagonal or triangular |
| eigenvalues with algebraic and geometric multiplicities | the row space, the column space |
| \( \operatorname{rank}(A - \lambda I)^k \) for every \( \lambda, k \) | orthogonality of the eigenvectors |
| Jordan form (up to block order) | the norm \( \lVert A \rVert \) |
Diagonalisation
For \( A \in M_n(F) \) the following are equivalent:
- \( A \) is diagonalisable over \( F \);
- \( \chi_A \) splits into linear factors over \( F \) and \( g(\lambda) = a(\lambda) \) for every eigenvalue \( \lambda \);
- \( \sum_\lambda \dim E_\lambda = n \);
- the minimal polynomial \( m_A \) splits over \( F \) with distinct roots.
Sufficient conditions to recognise at sight:
- \( n \) distinct eigenvalues in \( F \);
- \( p(A) = 0 \) for some \( p \) with distinct roots in \( F \): \( A^2 = A \), \( A^2 = I \) (\( \operatorname{char} F \ne 2 \)), \( A^3 = A \), \( A^k = I \) over \( \mathbb{C} \);
- real symmetric (orthogonally), Hermitian or more generally normal (unitarily); see Inner Product Spaces and Orthonormal Bases.
Signs of non-diagonalisability:
- a non-zero nilpotent matrix (only eigenvalue 0, but \( A \ne 0 \));
- \( A \ne \lambda I \) with a single eigenvalue \( \lambda \);
- \( \operatorname{rank}(A - \lambda I) > n - a(\lambda) \) for some eigenvalue.
Simultaneous diagonalisation. Two diagonalisable matrices are diagonalisable by the same \( P \) iff they commute. Commuting matrices over \( \mathbb{C} \) (diagonalisable or not) are simultaneously triangularisable.
Field matters. \( \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} \) is diagonalisable over \( \mathbb{C} \) (eigenvalues \( \pm i \), distinct) but not over \( \mathbb{R} \) (no real eigenvalues). Diagonalisable over \( \mathbb{R} \) needs all eigenvalues real.
Triangular Form
The diagonal of the triangular form lists the eigenvalues with algebraic multiplicity, which is why \( \operatorname{tr} A = \sum \lambda_i \) and \( \det A = \prod \lambda_i \) hold over \( \mathbb{C} \) for every matrix. Schur’s form is the standard route to the spectral theorem: a triangular normal matrix must be diagonal.
The Jordan Canonical Form
The Jordan block \( J_k(\lambda) \) is the \( k \times k \) matrix with \( \lambda \) on the diagonal, 1 on the superdiagonal and 0 elsewhere:
$$J_3(\lambda) = \begin{pmatrix} \lambda & 1 & 0 \\ 0 & \lambda & 1 \\ 0 & 0 & \lambda \end{pmatrix}.$$\( J_k(\lambda) - \lambda I \) is nilpotent of index exactly \( k \), with rank \( k - 1 \).
Reading the Jordan form
For each eigenvalue \( \lambda \) set \( r_k = \operatorname{rank}(A - \lambda I)^k \), with \( r_0 = n \).
| Quantity | Formula |
|---|---|
| total size of the \( \lambda \)-blocks | algebraic multiplicity \( a(\lambda) \) |
| number of \( \lambda \)-blocks | geometric multiplicity \( g(\lambda) = n - r_1 \) |
| largest \( \lambda \)-block | multiplicity of \( \lambda \) in the minimal polynomial |
| number of blocks of size \( \ge k \) | \( r_{k-1} - r_k \) |
| number of blocks of size exactly \( k \) | \( r_{k-1} - 2r_k + r_{k+1} \) |
Counting Jordan forms
For a given \( \chi_A = \prod (x - \lambda_i)^{a_i} \) and \( m_A = \prod (x - \lambda_i)^{m_i} \), the \( \lambda_i \)-blocks form a partition of \( a_i \) whose largest part is \( m_i \). The number of possible Jordan forms is the product, over the eigenvalues, of the number of such partitions.
| \( n \) | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
|---|---|---|---|---|---|---|---|
| partitions \( p(n) \) | 1 | 2 | 3 | 5 | 7 | 11 | 15 |
- Similarity classes of \( n \times n \) nilpotent matrices over \( \mathbb{C} \): \( p(n) \). For \( n = 4 \): \( 4, \ 3{+}1, \ 2{+}2, \ 2{+}1{+}1, \ 1{+}1{+}1{+}1 \).
- With \( \chi_A = (x - 1)^4(x + 2)^2 \) and \( m_A = (x - 1)^2(x + 2) \): for 1, partitions of 4 with largest part 2 are \( 2{+}2 \) and \( 2{+}1{+}1 \); for \( -2 \), partitions of 2 with largest part 1: only \( 1{+}1 \). Total \( 2 \times 1 = 2 \) Jordan forms.
Facts that follow from the Jordan form
- \( A \) is similar to \( A^T \) (each Jordan block is similar to its transpose by the reversal permutation).
- \( A \) is diagonalisable iff every Jordan block is \( 1 \times 1 \).
- Powers of a block: \( J_k(\lambda)^m \) is upper triangular with \( \binom{m}{j}\lambda^{m-j} \) on the \( j \)-th superdiagonal. For \( k = 2 \): \( \begin{pmatrix} \lambda^m & m\lambda^{m-1} \\ 0 & \lambda^m \end{pmatrix} \).
- \( A^m \to 0 \) as \( m \to \infty \) iff every eigenvalue has \( \lvert \lambda \rvert < 1 \).
- Every \( A \in M_n(\mathbb{C}) \) is \( D + N \) with \( D \) diagonalisable, \( N \) nilpotent and \( DN = ND \) (Jordan–Chevalley decomposition).
- Similarity over a field \( F \) is decided by invariants that do not change when \( F \) is enlarged (the rational canonical form, not separately in the syllabus). So two real matrices that are similar over \( \mathbb{C} \) are already similar over \( \mathbb{R} \).
graph TD
A["Square matrix A"] --> B{"chi_A splits over F?"}
B -->|no| C["Not triangularisable over F
(not diagonalisable either)"]
B -->|yes| D["Triangularisable; Jordan form exists"]
D --> E{"m_A has distinct roots?
(equivalently g = a for all eigenvalues)"}
E -->|yes| F["Diagonalisable: all blocks 1 x 1"]
E -->|no| G["Jordan blocks from ranks of (A - lambda I)^k"]Memory Tricks
- New basis in the columns. Put the new basis vectors as the columns of \( P \); the new matrix is \( P^{-1}AP \).
- Geometric counts, minimal measures. Geometric multiplicity = how many blocks; exponent in the minimal polynomial = how long the longest block.
- Rank drops count blocks. \( r_{k-1} - r_k \) = number of blocks of size at least \( k \).
- Jordan forms are partitions. For each eigenvalue, partition the algebraic multiplicity with largest part fixed by the minimal polynomial.
- Three is safe, four is not. \( \chi \) and \( m \) decide similarity up to \( n = 3 \).
Don’t Confuse
| This | Is not this | The difference |
|---|---|---|
| \( P^{-1}AP \) (new basis in columns of \( P \)) | \( PAP^{-1} \) | the second uses the inverse change of basis |
| similar | row equivalent | row equivalence (\( EA \)) preserves rank only, not eigenvalues |
| similar | congruent (\( P^TAP \)) | congruence preserves the inertia of symmetric matrices, not eigenvalues; see Quadratic Forms |
| triangularisable | diagonalisable | every complex matrix is triangularisable; not every one is diagonalisable |
| diagonalisable over \( \mathbb{C} \) | diagonalisable over \( \mathbb{R} \) | the second needs real eigenvalues |
| same \( \chi_A \) and \( m_A \) | similar | enough only for \( n \le 3 \) |
| number of Jordan blocks | algebraic multiplicity | number = geometric multiplicity |
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
- How many Jordan blocks for \( \lambda \) does \( A \) have, in terms of a rank?
- Is \( \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \) diagonalisable?
- Is a \( 3 \times 3 \) real matrix with eigenvalues \( 1, 2, 3 \) diagonalisable over \( \mathbb{R} \)?
Answers
- \( n - \operatorname{rank}(A - \lambda I) \), the geometric multiplicity.
- No. Only eigenvalue 1, and \( A \ne I \).
- Yes. Three distinct real eigenvalues give three independent real eigenvectors.
Level 2: Part B Style
Practice Q1. The number of similarity classes of \( 4 \times 4 \) nilpotent complex matrices is:
(a) 4 (b) 5 (c) 6 (d) 8
Answer: (b)
Similarity classes of nilpotent matrices correspond to Jordan forms with eigenvalue 0, i.e. partitions of 4: \( 4, 3{+}1, 2{+}2, 2{+}1{+}1, 1{+}1{+}1{+}1 \). So \( p(4) = 5 \).
Practice Q2. Let \( A \in M_6(\mathbb{C}) \) have characteristic polynomial \( (x - 1)^4(x + 2)^2 \) and minimal polynomial \( (x - 1)^2(x + 2) \). The number of possible Jordan forms of \( A \) (up to block order) is:
(a) 1 (b) 2 (c) 3 (d) 4
Answer: (b)
Eigenvalue 1: partitions of 4 with largest part 2: \( 2{+}2 \), \( 2{+}1{+}1 \). Eigenvalue \( -2 \): partitions of 2 with largest part 1: \( 1{+}1 \). Total \( 2 \times 1 = 2 \).
Practice Q3. A nilpotent \( N \in M_5(\mathbb{C}) \) has \( \operatorname{rank} N = 3 \), \( \operatorname{rank} N^2 = 1 \) and \( N^3 = 0 \). Its Jordan form is:
(a) \( J_5(0) \) (b) \( J_4(0) \oplus J_1(0) \) (c) \( J_3(0) \oplus J_2(0) \) (d) \( J_3(0) \oplus J_1(0) \oplus J_1(0) \)
Answer: (c)
Blocks: \( 5 - 3 = 2 \) in all; size \( \ge 2 \): \( 3 - 1 = 2 \); size \( \ge 3 \): \( 1 - 0 = 1 \). Sizes 3 and 2. Check the others: (a) has rank 4, (b) rank 3 but \( N^3 \ne 0 \), (d) rank 2.
Practice Q4. \( T : \mathbb{R}^2 \to \mathbb{R}^2 \) has matrix \( \begin{pmatrix} 1 & 2 \\ 0 & 3 \end{pmatrix} \) in the standard basis. Its matrix in the ordered basis \( ((1, 0), (1, 1)) \) is:
(a) \( \begin{pmatrix} 1 & 0 \\ 0 & 3 \end{pmatrix} \) (b) \( \begin{pmatrix} 1 & 2 \\ 0 & 3 \end{pmatrix} \) (c) \( \begin{pmatrix} 1 & 3 \\ 0 & 3 \end{pmatrix} \) (d) \( \begin{pmatrix} 3 & 0 \\ 0 & 1 \end{pmatrix} \)
Answer: (a)
\( T(1, 0) = (1, 0) = 1 \cdot (1, 0) + 0 \cdot (1, 1) \); \( T(1, 1) = (3, 3) = 0 \cdot (1, 0) + 3 \cdot (1, 1) \). Columns \( (1, 0)^T \) and \( (0, 3)^T \). Option (c) is \( AP \), the trap of forgetting \( P^{-1} \); (d) has the right entries in the wrong order for this ordered basis.
Level 3: Part C Style
Practice Q5. Let \( A \in M_4(\mathbb{C}) \) with \( A^3 = 0 \) but \( A^2 \ne 0 \). Which statements are true?
(a) \( \operatorname{rank} A = 2 \). (b) \( A \) is similar to \( A^T \). (c) \( \dim \ker A = 2 \). (d) \( \operatorname{rank} A^2 = 2 \).
Answer: (a), (b) and (c)
\( A \) is nilpotent; \( A^2 \ne 0 = A^3 \) makes the largest block exactly 3. The blocks partition 4 with largest part 3: only \( 3{+}1 \). So \( A \sim J_3(0) \oplus J_1(0) \).
- (a) True. \( \operatorname{rank} = 2 + 0 \).
- (b) True. Every square matrix is similar to its transpose.
- (c) True. Two blocks, so the kernel has dimension 2 (or \( 4 - 2 \)).
- (d) False. \( J_3(0)^2 \) has rank 1 and \( J_1(0)^2 = 0 \), so \( \operatorname{rank} A^2 = 1 \).
Practice Q6. Let \( R = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} \). Which statements are true?
(a) \( R \) is diagonalisable over \( \mathbb{C} \). (b) \( R \) is similar over \( \mathbb{R} \) to an upper triangular real matrix. (c) \( R^4 = I \). (d) \( R \) is similar over \( \mathbb{R} \) to \( -R \).
Answer: (a), (c) and (d)
- (a) True. \( \chi_R = x^2 + 1 \) has distinct roots \( \pm i \).
- (b) False. A real triangular form would have real eigenvalues on its diagonal; \( \chi_R \) does not split over \( \mathbb{R} \).
- (c) True. \( R^2 = -I \), so \( R^4 = I \).
- (d) True. With \( P = \operatorname{diag}(1, -1) \), \( PRP^{-1} = \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix} = -R \). (Also: \( R \) and \( -R \) have the same characteristic polynomial \( x^2 + 1 \) with distinct roots, so they are similar over \( \mathbb{C} \), hence over \( \mathbb{R} \).)
Practice Q7. Let \( A \) be a real \( 3 \times 3 \) matrix with characteristic polynomial \( (x - 1)^2(x - 2) \). Which statements are true?
(a) \( A \) is diagonalisable if and only if \( \operatorname{rank}(A - I) = 1 \). (b) \( (A - I)^2(A - 2I) = 0 \). (c) \( A^2 - 3A + 2I = 0 \) for every such \( A \). (d) \( A \) is invertible.
Answer: (a), (b) and (d)
- (a) True. Eigenvalue 2 is simple, so it is fine. Eigenvalue 1 has \( a = 2 \); diagonalisable iff \( g(1) = 3 - \operatorname{rank}(A - I) = 2 \), i.e. rank 1.
- (b) True. Cayley–Hamilton.
- (c) False. \( x^2 - 3x + 2 = (x - 1)(x - 2) \) kills \( A \) only when \( A \) is diagonalisable. For \( A = J_2(1) \oplus (2) \) it fails: \( (A - I)(A - 2I) \ne 0 \).
- (d) True. \( \det A = 1 \cdot 1 \cdot 2 = 2 \ne 0 \).
Quick Revision Table
| Idea | Fact |
|---|---|
| Change of basis | new basis as columns of \( P \); \( [T]_{\text{new}} = P^{-1}[T]_{\text{old}}P \) |
| Similarity invariants | rank, trace, det, \( \chi \), \( m \), eigenvalues, Jordan form |
| Diagonalisable | \( g = a \) for all \( \lambda \) \( \iff m_A \) has distinct roots (in \( F \)) |
| Sufficient | distinct eigenvalues; \( p(A) = 0 \) with \( p \) square-free and split; normal |
| Triangularisable over \( F \) | \( \iff \chi_A \) splits over \( F \); Schur: unitarily over \( \mathbb{C} \) |
| Number of \( \lambda \)-blocks | \( n - \operatorname{rank}(A - \lambda I) \) |
| Largest \( \lambda \)-block | exponent of \( (x - \lambda) \) in \( m_A \) |
| Blocks of size \( \ge k \) | \( r_{k-1} - r_k \) |
| Count Jordan forms | product of partition counts with fixed largest part |
| Nilpotent classes | \( p(n) \): 1, 2, 3, 5, 7, 11, 15 |
| \( \chi \) and \( m \) decide similarity | only for \( n \le 3 \) |
| \( A \sim A^T \) | always |
Key Takeaways
What’s Next
- Inner Product Spaces and Orthonormal Bases: when the change of basis can be chosen orthogonal or unitary, and the spectral theorem.
- Quadratic Forms: the other way of changing basis, \( P^TAP \), and Sylvester’s law of inertia.
- Eigenvalues and the Cayley–Hamilton Theorem to revise minimal polynomials.
- Linear Algebra Quick Revision.