When an operator is diagonalizable
An operator has a diagonal matrix exactly when some basis is made of its eigenvectors. Can that be decided without hunting for eigenvectors at all?
Where you are 5 in Eigenvalues, proved
Why this exists
Diagonalization proved that an n \times n matrix is diagonalizable exactly when it has n independent eigenvectors, and decided each case by solving for eigenspaces. For an operator on polynomials or matrices, or a matrix with a parameter in it, that is clumsy, and it does not say what is missing when the answer is no. This lesson proves the operator version over any field and adds a test with no eigenvectors in it: T is diagonalizable exactly when its The minimal polynomial is a product of distinct linear factors. The test passes to invariant subspaces, which gives commuting diagonalizable operators a common eigenvector basis.
Definition
Throughout, V is a finite-dimensional vector space over a field \mathbb{F} and T \in \mathcal{L}(V). The operator T is diagonalizable if [T]_{\mathcal{B}} is diagonal for some basis \mathcal{B} of V. A matrix A \in \mathbb{F}^{n \times n} is diagonalizable if \mathbf{x} \mapsto A\mathbf{x} on \mathbb{F}^n is; by Result 5 of The matrix of a linear map this means A = PDP^{-1} with P invertible and D diagonal, all entries in \mathbb{F}. The field is part of the question (Example 5).
When two operators are in play, E_\lambda(T) = \ker(T - \lambda I). A polynomial splits into distinct linear factors over \mathbb{F} if it equals (z - \mu_1)\cdots(z - \mu_r) with \mu_1, \dots, \mu_r \in \mathbb{F} distinct; for r = 0 it is 1.
Intuition
Compare A = \begin{bmatrix} 3 & 1 \\ 0 & 2 \end{bmatrix} and N = \begin{bmatrix} 2 & 1 \\ 0 & 2 \end{bmatrix} on \mathbb{R}^2. A has the eigenlines E_3 = \operatorname{span}\{(1, 0)\} and E_2 = \operatorname{span}\{(1, -1)\}, and the split of a vector along them can be written with A alone. The separating polynomials of Eigenvectors for distinct eigenvalues are independent are p_3(z) = z - 2 and p_2(z) = 3 - z. They add to 1, so \mathbf{v} = p_3(A)\mathbf{v} + p_2(A)\mathbf{v}, and
because (A - 3I)(A - 2I) = \begin{bmatrix} 0 & 1 \\ 0 & -1 \end{bmatrix}\begin{bmatrix} 1 & 1 \\ 0 & 0 \end{bmatrix} = 0. So p_3(A)\mathbf{v} \in E_3, and likewise p_2(A)\mathbf{v} \in E_2. For \mathbf{v} = (1, 1) the pieces are \mathbf{u} = (2, 0) and \mathbf{w} = (-1, 1).
For N the move fails. Its only eigenvalue is 2 and N - 2I \neq 0, so no product of distinct factors z - \mu kills N: its least annihilator is (z - 2)^2, and N turns \mathbf{e}_2 toward its one eigenline. Result 2 proves that a repeated factor, or a factor with no zero in \mathbb{F}, is the only possible obstruction.
Results
V is finite-dimensional over \mathbb{F} and T \in \mathcal{L}(V). All four results hold over any field.
Result 1 (conditions equivalent to diagonalizability). Let \lambda_1, \dots, \lambda_m be the distinct eigenvalues of T (m = 0 if there are none). The following are equivalent.
- (a) T is diagonalizable.
- (b) V has a basis consisting of eigenvectors of T.
- (c) V = E_{\lambda_1} \oplus \dots \oplus E_{\lambda_m}.
- (d) \dim V = \dim E_{\lambda_1} + \dots + \dim E_{\lambda_m}.
In particular \dim V distinct eigenvalues are enough (Result 5 of Eigenvectors for distinct eigenvalues are independent).
Result 2 (the minimal polynomial test). T is diagonalizable if and only if m_T splits into distinct linear factors over \mathbb{F}. In that case m_T(z) = (z - \lambda_1)\cdots(z - \lambda_m), the product over the distinct eigenvalues.
Result 3 (restrictions). If T is diagonalizable and U is a subspace invariant under T, then T|_U is diagonalizable.
Result 4 (commuting operators). Let S, T \in \mathcal{L}(V) be diagonalizable. Then V has a basis of vectors that are eigenvectors of both S and T if and only if ST = TS.
Why it is true
Result 1. If V = \{\mathbf{0}\}, all four hold, with the empty basis and the empty sum.
(a) \Leftrightarrow (b). Column j of [T]_{\mathcal{B}} holds the coordinates of T\mathbf{b}_j (The matrix of a linear map), so [T]_{\mathcal{B}} = \operatorname{diag}(d_1, \dots, d_n) exactly when T\mathbf{b}_j = d_j\mathbf{b}_j for each j; basis vectors are nonzero, so each \mathbf{b}_j is an eigenvector.
(b) \Rightarrow (c). Each vector of an eigenvector basis lies in some E_{\lambda_i}, so the subspace E_{\lambda_1} + \dots + E_{\lambda_m} contains a basis of V (Result 1 of Sums and direct sums of subspaces) and equals V. The sum is direct by Result 3 of Eigenvectors for distinct eigenvalues are independent.
(c) \Rightarrow (d) is Result 5 of Sums and direct sums of subspaces.
(d) \Rightarrow (b). Join bases of E_{\lambda_1}, \dots, E_{\lambda_m}. The joined list is independent (Result 3 of Eigenvectors for distinct eigenvalues are independent) of length \dim V, so it is a basis (Result 5 of Bases and dimension, proved) of eigenvectors.
Result 2. If V = \{\mathbf{0}\}, T is diagonalizable and m_T = 1, the empty product. Let V \neq \{\mathbf{0}\}.
Suppose T is diagonalizable, with an eigenvector basis \mathcal{B}, and let q(z) = (z - \lambda_1)\cdots(z - \lambda_m). Polynomials in T commute (Result 5 of Invariant subspaces), so for \mathbf{b} \in \mathcal{B} with T\mathbf{b} = \lambda_k\mathbf{b} the factor T - \lambda_kI may act first, and q(T)\mathbf{b} = \mathbf{0}. A linear map that vanishes on a basis is 0, so q(T) = 0. By Result 2(b) of The minimal polynomial, m_T is a product of some of the factors of q, and by Result 3(a) there every \lambda_k is a zero of m_T, so m_T = q.
Conversely, suppose m_T(z) = (z - \mu_1)\cdots(z - \mu_r) with the \mu_k distinct; r \ge 1 since V \neq \{\mathbf{0}\}. By Result 3(a) of The minimal polynomial, \mu_1, \dots, \mu_r are exactly the eigenvalues of T. Let \displaystyle p_k(z) = \prod_{j \neq k} \frac{z - \mu_j}{\mu_k - \mu_j} be the separating polynomials. Since p_k(\mu_i) is 1 for i = k and 0 otherwise, p_1 + \dots + p_r - 1 vanishes at r distinct points and has degree at most r - 1, so it is zero (The remainder and factor theorems):
Hence \mathbf{v} = p_1(T)\mathbf{v} + \dots + p_r(T)\mathbf{v} for every \mathbf{v} \in V. Since (z - \mu_k)p_k(z) = c_km_T(z) with c_k = \prod_{j \neq k}(\mu_k - \mu_j)^{-1},
so p_k(T)\mathbf{v} \in E_{\mu_k}. Thus V = E_{\mu_1} + \dots + E_{\mu_r}, direct by Result 3 of Eigenvectors for distinct eigenvalues are independent, and Result 1(c) holds.
Result 3. For \mathbf{u} \in U, invariance gives p(T|_U)\mathbf{u} = p(T)\mathbf{u} for every polynomial p (proof of Result 5 of The minimal polynomial), so m_T(T|_U) = 0. By Result 2, m_T is a product of distinct linear factors, and by Result 2(b) of The minimal polynomial for T|_U, so is m_{T|_U}. Result 2 for T|_U finishes.
Result 4. Suppose ST = TS and let \lambda_1, \dots, \lambda_m be the distinct eigenvalues of T. If \mathbf{v} \in E_{\lambda_i}(T), then (T - \lambda_iI)S\mathbf{v} = S(T - \lambda_iI)\mathbf{v} = \mathbf{0}, so S maps E_{\lambda_i}(T) into itself. By Results 3 and 1, E_{\lambda_i}(T) has a basis of eigenvectors of S, which are eigenvectors of T too. Since V = E_{\lambda_1}(T) \oplus \dots \oplus E_{\lambda_m}(T) (Result 1), the joined bases form a basis of V (Result 5 of Sums and direct sums of subspaces).
Conversely, in a basis \mathcal{B} of common eigenvectors, [S]_{\mathcal{B}} and [T]_{\mathcal{B}} are diagonal and so commute, and [ST]_{\mathcal{B}} = [S]_{\mathcal{B}}[T]_{\mathcal{B}} = [T]_{\mathcal{B}}[S]_{\mathcal{B}} = [TS]_{\mathcal{B}} (The matrix of a linear map), so ST = TS.
Worked examples
One product decides
Decide whether these matrices are diagonalizable:
Row 2 of K - \lambda I and of L - \lambda I is (0,\ 2 - \lambda,\ 0), so expanding along it (The characteristic equation and finding eigenvalues) gives, for both,
Both have exactly the eigenvalues 2 and 5. By Result 2, a diagonalizable one has minimal polynomial (z - 2)(z - 5), so that one product is the whole test. For K,
and the row (1, 1, 2) has dot product 0 with every column of K - 5I, so (K - 2I)(K - 5I) = 0 and K is diagonalizable. For L,
so L is not. No eigenvector was needed. As a check, E_5(K) = \operatorname{span}\{(1, 0, 1)\} and E_2(K) = \{\mathbf{x} : x_1 + x_2 + 2x_3 = 0\} have dimensions adding to 3, while L - 2I has rank 2, so \dim E_2(L) = 1.
A matrix with two parameters
For which a, b \in \mathbb{R} is A = \begin{bmatrix} 2 & a & 1 \\ 0 & 2 & b \\ 0 & 0 & 3 \end{bmatrix} diagonalizable?
SolutionA is upper triangular, so its eigenvalues are its diagonal entries 2 and 3 (Result 4 of Upper-triangular matrices over the complex numbers), and by Result 2 the test is whether (A - 2I)(A - 3I) = 0. The factors are
and their product is
This is 0 exactly when a = 0, whatever b is. For a = 0, E_2 = \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_2\} and E_3 = \operatorname{span}\{(1, b, 1)\}: the entry b is absorbed by the eigenvector for 3, while a ties two directions inside the one eigenvalue 2.
Every projection is diagonalizable
Let P \in \mathcal{L}(V) satisfy P^2 = P. Prove that P is diagonalizable, that \operatorname{im} P = E_1, and that V = \ker P \oplus \operatorname{im} P.
Solutionz(z - 1) annihilates P, so by Result 2(b) of The minimal polynomial m_P is one of 1, z, z - 1, z(z - 1). Each splits into distinct linear factors, so P is diagonalizable by Result 2, with eigenvalues among 0 and 1.
If \mathbf{v} = P\mathbf{u}, then P\mathbf{v} = P^2\mathbf{u} = \mathbf{v}; if P\mathbf{v} = \mathbf{v}, then \mathbf{v} \in \operatorname{im} P. So \operatorname{im} P = E_1, and \ker P = E_0 by definition. Result 1(c) gives V = \ker P \oplus \operatorname{im} P (a space E_0 or E_1 that is \{\mathbf{0}\} drops out). The pieces of \mathbf{v} are \mathbf{v} - P\mathbf{v} and P\mathbf{v}.
A reflection of cubics
Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), (Tp)(x) = p(2 - x). Show that T is diagonalizable, find a basis of each eigenspace, and split x^3 into its pieces.
Solution(T^2p)(x) = p(2 - (2 - x)) = p(x), so T^2 = I and (z - 1)(z + 1) annihilates T. As in Example 3, T is diagonalizable. Since T1 = 1 and Tx = 2 - x, T \neq \pm I, so m_T(z) = z^2 - 1 and both \pm 1 are eigenvalues.
In powers of t = x - 1 the eigenvectors are plain: replacing x by 2 - x sends t to -t, so T sends t^k to (-1)^kt^k. These four powers form a basis, and since the eigenspace dimensions add to 4 (Result 1(d)),
The separating polynomials for 1 and -1 are \tfrac12(z + 1) and -\tfrac12(z - 1), so the pieces of p are \tfrac12(p + Tp) and \tfrac12(p - Tp). With (2 - x)^3 = 8 - 12x + 6x^2 - x^3,
Check: in powers of t these are 3t^2 + 1 and t^3 + 3t, which lie in E_1 and E_{-1} and add to (t + 1)^3 = x^3.
The ringers’ shuffle
Yuna and Kofi ring eight bells hung in a belfry. A practice shuffle of their own moves the bell in place 2 to place 3, 3 to 5, 5 to 2, 4 to 7, 7 to 6 and 6 to 4, and leaves places 1 and 8 alone. Kofi records a number for each place as \mathbf{x} \in \mathbb{F}^8; the shuffle acts by the operator R with R\mathbf{e}_i = \mathbf{e}_{\sigma(i)}, where \sigma(i) is the new place of the bell from place i. Is R diagonalizable over \mathbb{C}? Over \mathbb{R}?
Solution\sigma cycles 2 \to 3 \to 5 \to 2 and 4 \to 7 \to 6 \to 4, so three shuffles bring every bell home: R^3 = I. Over \mathbb{C},
where \displaystyle \omega = -\tfrac12 + \tfrac{\sqrt{3}}{2}i, three distinct factors (Complex numbers). By Result 2(b) of The minimal polynomial, m_R is a product of some of them, so R is diagonalizable over \mathbb{C} (Result 2).
Over \mathbb{R} it is not. A real eigenvalue \lambda with eigenvector \mathbf{v} has \mathbf{v} = R^3\mathbf{v} = \lambda^3\mathbf{v}, so \lambda = 1, and a real diagonalization would give \mathbb{R}^8 = E_1 by Result 1(c), that is R = I, which is false.
Over \mathbb{C}, R\mathbf{x} = \mathbf{x} says the entries are constant along each cycle, leaving x_1, x_8 and one value per cycle free, so \dim E_1 = 4. Since \omega^2 = \bar{\omega}, the vector \mathbf{e}_2 + \bar{\omega}\mathbf{e}_3 + \omega\mathbf{e}_5 is an eigenvector for \omega, the other cycle gives a second one, and conjugating gives two for \bar{\omega}. The dimensions add to 8, so they are exactly 4, 2, 2. The E_1-piece of \mathbf{x} is \tfrac13(\mathbf{x} + R\mathbf{x} + R^2\mathbf{x}), from the separating polynomial \displaystyle \frac{(z - \omega)(z - \bar{\omega})}{(1 - \omega)(1 - \bar{\omega})} = \frac{z^2 + z + 1}{3}.
A common eigenvector basis
Let K be as in Example 1 and S = \begin{bmatrix} 1 & 0 & 0 \\ 2 & 3 & -2 \\ -1 & -1 & 2 \end{bmatrix}. Show that KS = SK and that S is diagonalizable, and find a basis of \mathbb{R}^3 of eigenvectors of both.
SolutionMultiplying, KS and SK both equal \begin{bmatrix} 3 & 1 & 2 \\ 4 & 6 & -4 \\ -1 & -1 & 6 \end{bmatrix}. Also (S - I)(S - 4I) = 0 while S is not a multiple of I, so m_S(z) = (z - 1)(z - 4) and S is diagonalizable (Result 2). By Result 4 a common basis exists, and its proof says where to look: inside each eigenspace of K.
E_5(K) = \operatorname{span}\{(1, 0, 1)\}, and S(1, 0, 1) = (1, 0, 1). E_2(K) is the plane x_1 + x_2 + 2x_3 = 0. S - I has every row a multiple of (1, 1, -1), so E_1(S) is the plane x_1 + x_2 - x_3 = 0, and the two planes meet along (1, -1, 0). A second eigenvector of S in E_2(K) is (0, 2, -1), with S(0, 2, -1) = (0, 8, -4). So
Boundaries
- The field matters. The shuffle R of Example 5 is diagonalizable over \mathbb{C} and not over \mathbb{R}, and so is the quarter turn (x, y) \mapsto (-y, x), with minimal polynomial z^2 + 1. Result 2 needs distinct factors that are linear over \mathbb{F}.
- Sums and products. On \mathbb{R}^2, S = \begin{bmatrix} 1 & 1 \\ 0 & 0 \end{bmatrix} is a projection and T = \operatorname{diag}(0, 1) is diagonal, but S + T = \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix} and ST = \begin{bmatrix} 0 & 1 \\ 0 & 0 \end{bmatrix} are not diagonalizable. These two do not commute; commuting ones behave (Exercise 16).
- Finite dimension. Result 2 needs m_T, and in infinite dimension an operator may have no nonzero annihilating polynomial at all. Condition (c) still makes sense in general and can fail: differentiation on \mathcal{P}(\mathbb{R}) is annihilated by no nonzero polynomial, has the single eigenvalue 0, and E_0 is only the constants.
Where this goes next
The spectral theorem finds, for self-adjoint and normal operators on an inner product space, an eigenvector basis that is also orthonormal. When m_T has a repeated factor, Generalized eigenvectors replaces each eigenspace by a larger invariant subspace so that V is still their direct sum, and Jordan canonical form reads the exponent of z - \lambda in m_T as the size of the largest block for \lambda.
Exercises
Testing for diagonalizability
- Find the minimal polynomial of A = \begin{bmatrix} 3 & -1 & 3 \\ 4 & -1 & 6 \\ 0 & 0 & 1 \end{bmatrix} and decide whether A is diagonalizable.
- Use Result 2 to show that B = \begin{bmatrix} 4 & 0 & 1 \\ 1 & 3 & 1 \\ 1 & 0 & 4 \end{bmatrix} is diagonalizable, then find a basis of each eigenspace.
- Show that C_c = \begin{bmatrix} 1 & 1 & 0 \\ 0 & c & 1 \\ 0 & 0 & 1 \end{bmatrix} is diagonalizable for no c \in \mathbb{R}.
- Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), (Tp)(x) = (x + 1)p'(x). Show that T is diagonalizable, give an eigenvector basis, and write x^2 in it.
- Let M = \begin{bmatrix} 2 & 1 \\ 0 & 2 \end{bmatrix} and TX = MX on \mathbb{R}^{2 \times 2}. Find m_T and \dim E_2, and decide whether T is diagonalizable.
- Let W = \operatorname{span}\{e^{3x}, xe^{3x}, e^{-x}\} and Df = f' on W. Find the matrix of D in this basis and m_D, and decide whether D is diagonalizable.
Using the test
- (a) Let V be a complex vector space and T^4 = I. Show that T is diagonalizable and list its possible eigenvalues. (b) Show that Q(x, y) = (-y, x) on \mathbb{R}^2 has Q^4 = I and is not diagonalizable.
- A real operator T satisfies T^3 = 4T. Show that T is diagonalizable and list its possible eigenvalues. Then check that A = \begin{bmatrix} 1 & 3 \\ 1 & -1 \end{bmatrix} satisfies A^3 = 4A, and find its eigenspaces.
- Kofi’s practice set has six bells. Its shuffle moves the bell in place 2 to 3, 3 to 5, 5 to 4 and 4 to 2, and leaves places 1 and 6 alone; R acts on \mathbb{F}^6 as in Example 5. Find the least k \ge 1 with R^k = I, decide whether R is diagonalizable over \mathbb{C} and over \mathbb{R}, and give \dim E_\lambda for each complex eigenvalue.
Proofs
- Suppose T^2 - 2T + I = 0. Prove that T is diagonalizable if and only if T = I.
- Let V be a complex vector space. Prove that T is diagonalizable if and only if \ker(T - \lambda I)^2 = \ker(T - \lambda I) for every \lambda \in \mathbb{C}.
- (a) Prove that if T is diagonalizable, then V = \ker T \oplus \operatorname{im} T. (b) Refute the converse with a counterexample.
- (a) Prove that if T is diagonalizable, then p(T) is diagonalizable for every polynomial p. (b) Refute: if T^2 is diagonalizable, then T is.
- Let V = U \oplus W with U and W invariant under T. Prove that T is diagonalizable if and only if T|_U and T|_W are.
- Let M \in \mathbb{F}^{n \times n} and TX = MX on \mathbb{F}^{n \times n}. Prove that m_T = m_M, so T is diagonalizable exactly when M is.
- Let S and T be diagonalizable with ST = TS. Prove that S + T and ST are diagonalizable.
Answers
- A - I = \begin{bmatrix} 2 & -1 & 3 \\ 4 & -2 & 6 \\ 0 & 0 & 0 \end{bmatrix} \neq 0 has rows that are multiples of (2, -1, 3) and columns that are multiples of (1, 2, 0), and (2, -1, 3) \cdot (1, 2, 0) = 0, so (A - I)^2 = 0. By Result 2(b) of The minimal polynomial and A \neq I, m_A(z) = (z - 1)^2. The factor repeats, so A is not diagonalizable.
- B - 3I has every row equal to (1, 0, 1), which has dot product 0 with every column of B - 5I = \begin{bmatrix} -1 & 0 & 1 \\ 1 & -2 & 1 \\ 1 & 0 & -1 \end{bmatrix}. So (B - 3I)(B - 5I) = 0 and B is diagonalizable. E_3 = \{\mathbf{x} : x_1 + x_3 = 0\} has basis (1, 0, -1), (0, 1, 0), and E_5 = \operatorname{span}\{(1, 1, 1)\}.
- C_c is upper triangular with eigenvalues 1 and c (Result 4 of Upper-triangular matrices over the complex numbers). For c \neq 1, Result 2 asks for (C_c - I)(C_c - cI) = 0, but its (1, 3) entry is row (0, 1, 0) of C_c - I against column (0, 1, 1 - c) of C_c - cI, which is 1. For c = 1 it asks for C_1 - I = 0, which is false.
- T(x + 1)^k = (x + 1)\,k(x + 1)^{k - 1} = k(x + 1)^k, so 1, x + 1, (x + 1)^2, (x + 1)^3 are eigenvectors for 0, 1, 2, 3: four distinct eigenvalues on a 4-dimensional space, so they form a basis (Result 5 of Eigenvectors for distinct eigenvalues are independent). In it, x^2 = (x + 1)^2 - 2(x + 1) + 1.
- (T - 2I)X = NX with N = M - 2I = \begin{bmatrix} 0 & 1 \\ 0 & 0 \end{bmatrix}, and N^2 = 0 \neq N, so (T - 2I)^2 = 0 \neq T - 2I and m_T(z) = (z - 2)^2: not diagonalizable. NX = 0 exactly when the second row of X is zero, so \dim E_2 = 2 < 4.
- De^{3x} = 3e^{3x}, D(xe^{3x}) = e^{3x} + 3xe^{3x} and De^{-x} = -e^{-x}, so the matrix is \begin{bmatrix} 3 & 1 & 0 \\ 0 & 3 & 0 \\ 0 & 0 & -1 \end{bmatrix}. By Result 6(a) of The minimal polynomial, (D - 3I)^2(D + I) = 0, and 3, -1 are zeros of m_D. But (D + I)(xe^{3x}) = e^{3x} + 4xe^{3x}, and D - 3I sends that to 4e^{3x} \neq 0, so m_D(z) = (z - 3)^2(z + 1): not diagonalizable.
- (a) z^4 - 1 = (z - 1)(z + 1)(z - i)(z + i) has distinct factors, and m_T is a product of some of them, so T is diagonalizable with eigenvalues among 1, -1, i, -i. (b) Q^2(x, y) = Q(-y, x) = (-x, -y), so Q^2 = -I and Q^4 = I. A real eigenvalue \lambda would give -\mathbf{v} = \lambda^2\mathbf{v} with \mathbf{v} \neq \mathbf{0}, impossible, so Result 1(d) fails: 0 \neq 2.
- z^3 - 4z = z(z - 2)(z + 2) has distinct real factors, so T is diagonalizable with eigenvalues among 0, 2, -2. A^2 = 4I, so A^3 = 4A; as A \neq \pm 2I, m_A(z) = z^2 - 4 and both \pm 2 are eigenvalues: E_2 = \operatorname{span}\{(3, 1)\} and E_{-2} = \operatorname{span}\{(1, -1)\}.
- Places 2, 3, 5, 4 form one cycle of length 4, so k = 4, and by Exercise 7(a) R is diagonalizable over \mathbb{C}. E_1: entries constant on the cycle, x_1 and x_6 free, dimension 3. E_{-1} is spanned by \mathbf{e}_2 - \mathbf{e}_3 + \mathbf{e}_5 - \mathbf{e}_4; E_i contains \mathbf{e}_2 - i\mathbf{e}_3 - \mathbf{e}_5 + i\mathbf{e}_4 and E_{-i} its conjugate. The dimensions add to 6, so they are 3, 1, 1, 1. Over \mathbb{R} only \pm 1 remain and 3 + 1 < 6: not diagonalizable.
- If T = I, its matrix is diagonal in every basis. Conversely, (z - 1)^2 annihilates T, so m_T is 1, z - 1 or (z - 1)^2. If T is diagonalizable, Result 2 excludes (z - 1)^2; m_T = z - 1 gives T = I, and m_T = 1 means V = \{\mathbf{0}\}, where T = I as well.
- Always \ker(T - \lambda I) \subseteq \ker(T - \lambda I)^2. Let T be diagonalizable and (T - \lambda I)^2\mathbf{v} = \mathbf{0}. Write \mathbf{v} = \mathbf{v}_1 + \dots + \mathbf{v}_m with \mathbf{v}_i \in E_{\lambda_i} (Result 1(c)). Then \mathbf{0} = \sum_i (\lambda_i - \lambda)^2\mathbf{v}_i, and directness gives (\lambda_i - \lambda)^2\mathbf{v}_i = \mathbf{0}, so \mathbf{v}_i = \mathbf{0} whenever \lambda_i \neq \lambda, and \mathbf{v} \in E_\lambda. Conversely, assume the kernels agree. If V = \{\mathbf{0}\}, T is diagonalizable with the empty basis, so let V \neq \{\mathbf{0}\}. By the The fundamental theorem of algebra, m_T is a product of linear factors. If one repeated, m_T = (z - \lambda)^2q, and by minimality (T - \lambda I)q(T)\mathbf{v} \neq \mathbf{0} for some \mathbf{v}. Then \mathbf{w} = q(T)\mathbf{v} has (T - \lambda I)^2\mathbf{w} = m_T(T)\mathbf{v} = \mathbf{0} but (T - \lambda I)\mathbf{w} \neq \mathbf{0}, a contradiction. So Result 2 applies.
- (a) Take an eigenvector basis with T\mathbf{b}_j = d_j\mathbf{b}_j; let \mathcal{B}_0 be the \mathbf{b}_j with d_j = 0 and \mathcal{B}_1 the rest. As T(\sum_j c_j\mathbf{b}_j) = \sum_j c_jd_j\mathbf{b}_j, \ker T = \operatorname{span}\mathcal{B}_0, and \operatorname{im} T = \operatorname{span}\mathcal{B}_1, since each \mathbf{b}_j \in \mathcal{B}_1 is T(d_j^{-1}\mathbf{b}_j). Two parts of one basis span a direct sum equal to V (Result 5 of Sums and direct sums of subspaces). (b) T(x, y) = (x + y, y) on \mathbb{R}^2 is invertible, so \ker T \oplus \operatorname{im} T = \{\mathbf{0}\} \oplus \mathbb{R}^2. But (T - I)(x, y) = (y, 0), so (T - I)^2 = 0 and T \neq I: by Exercise 10, T is not diagonalizable.
- (a) If T\mathbf{b} = d\mathbf{b}, then p(T)\mathbf{b} = p(d)\mathbf{b} (first step of the proof of Result 1 of Eigenvectors for distinct eigenvalues are independent), so an eigenvector basis of T is one of p(T). (b) N(x, y) = (y, 0) has N^2 = 0, which is diagonal, but N \neq 0, so m_N(z) = z^2 repeats a factor.
- If T is diagonalizable, Result 3 gives both restrictions. Conversely, an eigenvector basis of U for T|_U and one of W for T|_W, joined, form a basis of V (Result 5 of Sums and direct sums of subspaces) of eigenvectors of T.
- T^kX = M^kX by induction (T^{k+1}X = M(M^kX)), so p(T)X = p(M)X for every polynomial p. If p(M) = 0, then p(T) = 0; if p(T) = 0, then p(M) = p(M)I_n = p(T)I_n = 0. So T and M have the same annihilating polynomials, hence the same minimal polynomial, and Result 2 decides both at once.
- By Result 4 there is a basis \mathcal{B} of common eigenvectors. Then [S + T]_{\mathcal{B}} = [S]_{\mathcal{B}} + [T]_{\mathcal{B}} and [ST]_{\mathcal{B}} = [S]_{\mathcal{B}}[T]_{\mathcal{B}} are a sum and a product of diagonal matrices, hence diagonal.