Eigenvectors for distinct eigenvalues are independent
Eigenvectors that belong to different eigenvalues never conspire to cancel. Why not, and how many different eigenvalues can one operator have?
Where you are 5 in Eigenvalues, proved
Why this exists
Finding an eigenvector proved, by induction on the number of vectors, that eigenvectors of an n \times n matrix for distinct eigenvalues are linearly independent, and Diagonalization used it. This lesson proves the same fact for an operator T \in \mathcal{L}(V) on any vector space over any field, including spaces of functions and sequences with no finite basis, and by a different argument: for each eigenvalue there is a polynomial in T that keeps the piece of a vector belonging to that eigenvalue and erases every other piece.
Two consequences carry the rest of the unit. The eigenspaces of T always form a direct sum, so their dimensions add up to at most \dim V; in particular an operator on an n-dimensional space has at most n eigenvalues. And n distinct eigenvalues on an n-dimensional space hand over a basis of eigenvectors, in which the matrix of T is diagonal.
Definition
Throughout, V is a vector space over a field \mathbb{F} and T \in \mathcal{L}(V). As in Invariant subspaces, \lambda \in \mathbb{F} is an eigenvalue of T if T\mathbf{v} = \lambda\mathbf{v} for some \mathbf{v} \neq \mathbf{0}, such a \mathbf{v} is an eigenvector for \lambda, and the eigenspace is E_\lambda = \ker(T - \lambda I). When V is finite-dimensional, \dim E_\lambda is the geometric multiplicity of \lambda.
Eigenvalues \lambda_1, \dots, \lambda_m are distinct if \lambda_i \neq \lambda_j whenever i \neq j. For distinct scalars \lambda_1, \dots, \lambda_m and each k, the separating polynomial for \lambda_k is
the product over every j from 1 to m except k. The denominators are nonzero because the scalars are distinct. p_k has degree m - 1, with p_k(\lambda_k) = 1 and p_k(\lambda_i) = 0 for i \neq k. For m = 1 the product is empty and p_1 = 1.
Intuition
T stretches an eigenvector for \lambda by the factor \lambda. Take \mathbf{v} = \mathbf{v}_1 + \mathbf{v}_2 with T\mathbf{v}_1 = \lambda_1\mathbf{v}_1 and T\mathbf{v}_2 = \lambda_2\mathbf{v}_2. The operator T - \lambda_2 I sends \mathbf{v}_2 to \mathbf{0} and multiplies \mathbf{v}_1 by \lambda_1 - \lambda_2:
If \lambda_1 \neq \lambda_2, dividing by \lambda_1 - \lambda_2 recovers \mathbf{v}_1 from \mathbf{v} alone. So if \mathbf{v}_1 + \mathbf{v}_2 were \mathbf{0}, then \mathbf{v}_1 would be \mathbf{0} too: pieces stretched by different factors cannot cancel. With m pieces, one factor T - \lambda_j I for each unwanted eigenvalue erases them all, and that product, scaled, is p_k(T).
The count follows from room. Independent eigenvectors take up dimensions, and an n-dimensional space holds at most n independent vectors.
Results
Results 1 to 3 hold for every vector space V over every field \mathbb{F}, finite-dimensional or not. Results 4 and 5 need V finite-dimensional.
Result 1 (separating the pieces). Let \lambda_1, \dots, \lambda_m be distinct scalars and \mathbf{v} = \mathbf{v}_1 + \dots + \mathbf{v}_m with each \mathbf{v}_i \in E_{\lambda_i}. Then p_k(T)\mathbf{v} = \mathbf{v}_k for each k.
Result 2 (eigenvectors for distinct eigenvalues are independent). If \lambda_1, \dots, \lambda_m are distinct eigenvalues of T and \mathbf{v}_i is an eigenvector for \lambda_i, then (\mathbf{v}_1, \dots, \mathbf{v}_m) is linearly independent.
Result 3 (eigenspaces form a direct sum). If \lambda_1, \dots, \lambda_m are distinct eigenvalues of T, the sum E_{\lambda_1} + \dots + E_{\lambda_m} is direct. If each E_{\lambda_i} is finite-dimensional, bases of them joined in order form an independent list.
Result 4 (at most \dim V eigenvalues). Let \dim V = n. If \lambda_1, \dots, \lambda_m are distinct eigenvalues of T, then
In particular T has at most n distinct eigenvalues.
Result 5 (n distinct eigenvalues). Let \dim V = n \ge 1 and let T have n distinct eigenvalues \lambda_1, \dots, \lambda_n with eigenvectors \mathbf{v}_1, \dots, \mathbf{v}_n. Then T has no other eigenvalue, each E_{\lambda_i} = \operatorname{span}\{\mathbf{v}_i\}, V = E_{\lambda_1} \oplus \dots \oplus E_{\lambda_n}, and \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n) is a basis of V with [T]_{\mathcal{B}} = \operatorname{diag}(\lambda_1, \dots, \lambda_n).
Why it is true
Result 1. First, if T\mathbf{u} = \lambda\mathbf{u}, then T^j\mathbf{u} = \lambda^j\mathbf{u} for every j \ge 0 by Proof by induction: T^0\mathbf{u} = \mathbf{u}, and T^{j+1}\mathbf{u} = T(\lambda^j\mathbf{u}) = \lambda^{j+1}\mathbf{u} by linearity. So for p(z) = \sum_j a_jz^j, p(T)\mathbf{u} = \sum_j a_j\lambda^j\mathbf{u} = p(\lambda)\mathbf{u}. This holds for \mathbf{u} = \mathbf{0} as well, so it holds for every \mathbf{u} \in E_\lambda. Now p_k(T) is linear, so
since p_k(\lambda_k) = 1 and p_k(\lambda_i) = 0 for i \neq k.
Result 2. Suppose c_1\mathbf{v}_1 + \dots + c_m\mathbf{v}_m = \mathbf{0}. Each c_i\mathbf{v}_i lies in E_{\lambda_i}, so Result 1 applied to this sum gives, for each k,
As \mathbf{v}_k \neq \mathbf{0}, c_k = 0. Every coefficient is 0, so the list is independent.
Result 3. By the zero test (Result 2 of Sums and direct sums of subspaces) it is enough to show that \mathbf{0} = \mathbf{u}_1 + \dots + \mathbf{u}_m with each \mathbf{u}_i \in E_{\lambda_i} forces every \mathbf{u}_i = \mathbf{0}. Result 1 gives \mathbf{u}_k = p_k(T)\mathbf{0} = \mathbf{0} for each k. The joined bases are independent by Result 5 of Sums and direct sums of subspaces.
Result 4. Each E_{\lambda_i} is a subspace of V, so it is finite-dimensional (Result 6 of Bases and dimension, proved). By Result 3 and Result 5 of Sums and direct sums of subspaces, the direct sum E_{\lambda_1} \oplus \dots \oplus E_{\lambda_m} has dimension \dim E_{\lambda_1} + \dots + \dim E_{\lambda_m}. It is a subspace of V (Result 1 there), so its dimension is at most n by Result 6 of Bases and dimension, proved. Each \lambda_i has an eigenvector, so \dim E_{\lambda_i} \ge 1, and then m \le \dim E_{\lambda_1} + \dots + \dim E_{\lambda_m} \le n.
Result 5. If \mu were an eigenvalue different from all \lambda_i, then \lambda_1, \dots, \lambda_n, \mu would be n + 1 distinct eigenvalues, against Result 4. Result 4 also gives \sum_i \dim E_{\lambda_i} \le n with each term at least 1, so every term is 1; as \mathbf{v}_i \in E_{\lambda_i} is nonzero, E_{\lambda_i} = \operatorname{span}\{\mathbf{v}_i\}. By Result 2, \mathcal{B} is an independent list of length n, so it is a basis (Result 5 of Bases and dimension, proved); it spans V, so V = E_{\lambda_1} + \dots + E_{\lambda_n}, a direct sum by Result 3. Column j of [T]_{\mathcal{B}} holds the coordinates of T\mathbf{v}_j = \lambda_j\mathbf{v}_j (The matrix of a linear map), which are \lambda_j in place j and 0 elsewhere.
Worked examples
Three eigenvectors and the pieces of a vector
Let A = \begin{bmatrix} -1 & -3 & 3 \\ -4 & -1 & 4 \\ -4 & -3 & 6 \end{bmatrix} act on \mathbb{R}^3. Show that (1, 0, 1), (1, 1, 1) and (0, 1, 1) are eigenvectors, conclude that they form a basis, and split \mathbf{w} = (3, 4, 6) into its pieces in the eigenspaces using only A.
SolutionMultiplying out,
The eigenvalues 2, -1 and 3 are distinct, so by Result 5 the three vectors form a basis of \mathbb{R}^3, A has no other eigenvalue, and each eigenspace is a line. No row reduction was needed.
By Result 1, the piece of \mathbf{w} in E_2 is p_1(A)\mathbf{w} with \displaystyle p_1(z) = \frac{(z + 1)(z - 3)}{(2 + 1)(2 - 3)}. First A\mathbf{w} = (3, 8, 12), so (A - 3I)\mathbf{w} = (-6, -4, -6). Then A(-6, -4, -6) = (0, 4, 0), so (A + I)(-6, -4, -6) = (-6, 0, -6), and dividing by (3)(-1) = -3 gives the piece (2, 0, 2).
For E_3, \displaystyle p_3(z) = \frac{(z - 2)(z + 1)}{(3 - 2)(3 + 1)}. Here (A + I)\mathbf{w} = (6, 12, 18) and A(6, 12, 18) = (12, 36, 48), so (A - 2I)(6, 12, 18) = (0, 12, 12) and the piece is (0, 3, 3). The pieces add to \mathbf{w}, so the piece in E_{-1} is
So \mathbf{w} = 2(1, 0, 1) + (1, 1, 1) + 3(0, 1, 1), found without solving a system.
Exponentials under differentiation
Let C^{\infty}(\mathbb{R}) be the real functions on \mathbb{R} with derivatives of every order, and Df = f'. Prove that for distinct real numbers \lambda_1, \dots, \lambda_m the list (e^{\lambda_1x}, \dots, e^{\lambda_mx}) is linearly independent. Then recover the e^{2x} term of f = 4e^{2x} - e^{-x} + 3e^{x/2} using only derivatives.
SolutionC^{\infty}(\mathbb{R}) is the intersection of the subspaces C^{k}(\mathbb{R}) of Polynomials and functions as vectors, so it is a subspace, and D maps it to itself linearly. The chain rule gives De^{\lambda x} = \lambda e^{\lambda x}, and e^{\lambda x} is not the zero function, so it is an eigenvector of D for \lambda. The eigenvalues \lambda_1, \dots, \lambda_m are distinct, so Result 2 gives independence. Result 2 needs no finite dimension. Taking \lambda_j = j for every m shows again that C^{\infty}(\mathbb{R}) is infinite-dimensional (Result 7 of Span and independence, proved), and here any distinct exponents work, not only integers.
The terms of f lie in E_2, E_{-1} and E_{1/2}. By Result 1, 4e^{2x} = p(D)f with
Apply the factors one at a time. With f' = 8e^{2x} + e^{-x} + \tfrac32e^{x/2},
and then
Multiplying by \tfrac29 gives 4e^{2x}, the term we wanted.
All eigenvalues of an operator on matrices
Let M = \begin{bmatrix} 1 & 2 \\ 0 & 3 \end{bmatrix} and T \in \mathcal{L}(\mathbb{R}^{2 \times 2}), TX = MX - XM. Find every eigenvalue of T and a basis of each eigenspace.
SolutionIf M\mathbf{u} = \mu\mathbf{u} and \mathbf{w}^{T}M = \nu\mathbf{w}^{T}, then T(\mathbf{u}\mathbf{w}^{T}) = \mu\mathbf{u}\mathbf{w}^{T} - \nu\mathbf{u}\mathbf{w}^{T}. The matrix M has eigenvectors (1, 0) for 1 and (1, 1) for 3; the rows (1, -1) and (0, 1) satisfy (1, -1)M = (1, -1) and (0, 1)M = 3(0, 1). The four products are
with TX_1 = TX_2 = 0, TX_3 = -2X_3 (from \mu = 1, \nu = 3) and TX_4 = 2X_4 (from \mu = 3, \nu = 1). Direct multiplication confirms each, for instance
So 0, 2 and -2 are eigenvalues. X_1 and X_2 are not multiples of each other, so \dim E_0 \ge 2, while \dim E_2 \ge 1 and \dim E_{-2} \ge 1. By Result 4 the three dimensions add to at most \dim \mathbb{R}^{2 \times 2} = 4, so they are exactly 2, 1, 1, with bases (X_1, X_2), (X_4) and (X_3). A fourth eigenvalue would add at least 1 more and break the bound, so there is none. By Result 3, \mathbb{R}^{2 \times 2} = E_0 \oplus E_2 \oplus E_{-2}. As a check, I = X_1 + X_2 and M = X_1 + 3X_2 lie in E_0, as they must, since both commute with M.
Operators that commute with T
Let \dim V = n and let T \in \mathcal{L}(V) have n distinct eigenvalues, with eigenvectors \mathbf{v}_1, \dots, \mathbf{v}_n. Prove that if S \in \mathcal{L}(V) and ST = TS, then each \mathbf{v}_k is an eigenvector of S, so [S]_{\mathcal{B}} is diagonal for \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n).
SolutionLet \lambda_k be the eigenvalue of \mathbf{v}_k. Since S commutes with T and with I, it commutes with T - \lambda_kI. So (T - \lambda_kI)S\mathbf{v}_k = S(T - \lambda_kI)\mathbf{v}_k = S\mathbf{0} = \mathbf{0}, and S\mathbf{v}_k \in E_{\lambda_k}. By Result 5, E_{\lambda_k} = \operatorname{span}\{\mathbf{v}_k\}, so S\mathbf{v}_k = \mu_k\mathbf{v}_k for some \mu_k \in \mathbb{F}, and \mathbf{v}_k \neq \mathbf{0} makes it an eigenvector of S. Column k of [S]_{\mathcal{B}} is then \mu_k in place k and 0 elsewhere: [S]_{\mathcal{B}} = \operatorname{diag}(\mu_1, \dots, \mu_n).
For the matrix A of Example 1, solving SA = AS for a general 3 \times 3 matrix S leaves three free entries, and every solution has (1, 0, 1), (1, 1, 1) and (0, 1, 1) as eigenvectors. Exercise 14 goes further: such an S is a polynomial in T.
Counting drum cadences
Samir writes cadences for a drumline. A cadence of k beats is a row of single taps, each one beat long, and two-beat figures, of which the section plays 6 kinds. Let a_k be the number of cadences of k beats, with a_0 = 1 (the empty cadence) and a_1 = 1. Sorting by the last element gives a_k = a_{k-1} + 6a_{k-2} for k \ge 2. Amara asks for a formula for a_k and the number of 10-beat cadences.
SolutionLet L be the left shift on the space \mathbb{R}^{\infty} of real sequences (a_0, a_1, a_2, \dots), so (La)_k = a_{k+1}; it is linear. Let W be the set of sequences with a_{k+2} = a_{k+1} + 6a_k for every k \ge 0. The condition is preserved by sums and scalar multiples, so W is a subspace, and it is invariant under L: if a \in W, then (La)_{k+2} = a_{k+3} = a_{k+2} + 6a_{k+1}, which is (La)_{k+1} + 6(La)_k.
The map a \mapsto (a_0, a_1) from W to \mathbb{R}^2 is linear. It is injective, since the recurrence fixes every later term from the first two (by Strong induction and well-ordering), and surjective, since any pair of starting values extends by the recurrence to a member of W. So W \cong \mathbb{R}^2 and \dim W = 2.
A geometric sequence (r^k) with r \neq 0 lies in W when r^2 = r + 6, that is (r - 3)(r + 2) = 0. And L(r^k) = (r^{k+1}) = r(r^k), so (3^k) and ((-2)^k) are eigenvectors of L|_W for the distinct eigenvalues 3 and -2. By Result 5 on the 2-dimensional W, they form a basis, and a_k = c_13^k + c_2(-2)^k. The starting values give c_1 + c_2 = 1 and 3c_1 - 2c_2 = 1, so c_1 = \tfrac35 and c_2 = \tfrac25:
This gives a_2 = 7, a_3 = 13, a_4 = 55, matching the recurrence, and
Boundaries
- The converse fails. Independent eigenvectors may share an eigenvalue: every basis of \mathbb{F}^n consists of eigenvectors of I, all for 1.
- Distinct eigenvalues are enough for an eigenbasis, not needed. I has one eigenvalue and a basis of eigenvectors, while T(x, y) = (x + y, y) has the one eigenvalue 1 with E_1 = \operatorname{span}\{\mathbf{e}_1\} and no eigenbasis. When an operator is diagonalizable decides exactly when one exists.
- Results 4 and 5 need finite dimension. On \mathbb{F}^{\infty} every scalar is an eigenvalue of the left shift (Exercise 13), and on C^{\infty}(\mathbb{R}) every real number is an eigenvalue of D (Example 2).
- Eigenvalues are counted in \mathbb{F}. Over \mathbb{R} an operator may have fewer than n eigenvalues, even none (Invariant subspaces, Example 4), and then Result 5 does not apply.
Where this goes next
When an operator is diagonalizable uses Result 3 to show that T has a diagonal matrix exactly when V = E_{\lambda_1} \oplus \dots \oplus E_{\lambda_m}. The product (z - \lambda_1)\cdots(z - \lambda_m) hidden in the separating polynomials returns in The minimal polynomial, and Generalized eigenvectors proves the same independence for generalized eigenvectors. Upper-triangular matrices over the complex numbers finds an eigenvalue over \mathbb{C} when Result 5 is out of reach.
Exercises
Eigenvectors and their pieces
- Find the eigenvalues and eigenvectors of A = \begin{bmatrix} 4 & 1 \\ 2 & 3 \end{bmatrix}, and split \mathbf{w} = (7, 1) into its pieces in the eigenspaces with the separating polynomials.
- Let B = \begin{bmatrix} -2 & 3 & -3 \\ -12 & 13 & -9 \\ -12 & 12 & -8 \end{bmatrix}. Show that (1, 1, 0), (0, 1, 1) and (1, 2, 2) are eigenvectors of B, and give the matrix of B in the basis they form. Why is it a basis?
- For B of Exercise 2 and \mathbf{w} = (2, 5, 4), find the pieces of \mathbf{w} in E_4 and E_1 with the separating polynomials, then the piece in E_{-2}.
- Show that (e^{x}, e^{3x}, e^{-2x}) is independent, and for f = 2e^{x} + e^{3x} - 4e^{-2x} compute (D - 3I)(D + 2I)f by taking derivatives. Which term of f does it recover, after dividing by what?
- Over \mathbb{C}, let D act on the complex-valued functions on \mathbb{R} with derivatives of every order. Show that e^{ix} and e^{-ix} are eigenvectors of D, and deduce that (\cos x, \sin x) is independent.
- Let W be the real sequences with a_{k+2} = 3a_{k+1} + 4a_k for every k \ge 0. Find the two geometric sequences in W, show that they form a basis of W, and find the member with a_0 = 2 and a_1 = 3.
Counting eigenvalues
- Let M = \operatorname{diag}(2, -1) and TX = MX - XM on \mathbb{R}^{2 \times 2}. Find every eigenvalue of T and each eigenspace, and explain why there are no others.
- Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), (Tp)(x) = xp'(x) - p(x). Find four eigenvalues, explain without solving any equation why each eigenspace is a line and there are no others, and find \ker T and \operatorname{im} T.
- Let \dim V = 6, and let T \in \mathcal{L}(V) have eigenvalues 1, 2, 3 with \dim E_1 = 2 and \dim E_2 = 3. Find \dim E_3, decide whether T can have a fourth eigenvalue, and give [T]_{\mathcal{B}} for \mathcal{B} the joined bases of E_1, E_2, E_3.
- Can an operator on \mathbb{R}^5 have three eigenvalues whose eigenspaces all have dimension 2?
Proofs
- Let \mathbf{v}_1, \dots, \mathbf{v}_m, with m \ge 2, be eigenvectors of T for distinct eigenvalues. Prove that \mathbf{v}_1 + \dots + \mathbf{v}_m is not an eigenvector of T.
- Let V be finite-dimensional. Prove that T has at most \operatorname{rank} T + 1 distinct eigenvalues. How many can an operator on \mathbb{F}^5 of rank 1 have?
- Prove that if T has infinitely many distinct eigenvalues, V is infinite-dimensional, and show that every \lambda \in \mathbb{F} is an eigenvalue of the left shift on \mathbb{F}^{\infty}.
- Let \dim V = n and let T have n distinct eigenvalues. Prove that every S \in \mathcal{L}(V) with ST = TS equals p(T) for some polynomial p of degree less than n.
- Let \lambda_1, \dots, \lambda_m be distinct and \mathbf{v} = \mathbf{v}_1 + \dots + \mathbf{v}_m with \mathbf{v}_i \in E_{\lambda_i}. Prove that every \mathbf{v}_i lies in \operatorname{span}\{\mathbf{v}, T\mathbf{v}, \dots, T^{m-1}\mathbf{v}\}.
- Let U be invariant under T and \mathbf{v} as in Exercise 15 with \mathbf{v} \in U. Prove that every \mathbf{v}_i \in U. Deduce that if V = E_{\lambda_1} \oplus \dots \oplus E_{\lambda_m}, then U = (U \cap E_{\lambda_1}) \oplus \dots \oplus (U \cap E_{\lambda_m}).
- Refute with a counterexample: if T has \dim V distinct eigenvalues, so does T^2.
Answers
- \det(A - \lambda I) = \lambda^2 - 7\lambda + 10 = (\lambda - 5)(\lambda - 2). E_5 = \operatorname{span}\{(1, 1)\} and E_2 = \operatorname{span}\{(1, -2)\}. The piece in E_5 is \tfrac13(A - 2I)\mathbf{w}, that is \tfrac13(15, 15) = (5, 5); the piece in E_2 is -\tfrac13(A - 5I)\mathbf{w}, that is -\tfrac13(-6, 12) = (2, -4). Check: (5, 5) + (2, -4) = (7, 1).
- B(1, 1, 0) = (1, 1, 0), B(0, 1, 1) = (0, 4, 4) and B(1, 2, 2) = (-2, -4, -4), for the eigenvalues 1, 4, -2. They are distinct, so by Result 5 the three vectors are a basis of \mathbb{R}^3, and the matrix is \operatorname{diag}(1, 4, -2).
- B\mathbf{w} = (-1, 5, 4), so (B + 2I)\mathbf{w} = (3, 15, 12) and B(3, 15, 12) = (3, 51, 48). Piece in E_4: (B - I)(3, 15, 12) = (0, 36, 36), and dividing by (4 - 1)(4 + 2) = 18 gives (0, 2, 2). Piece in E_1: (B - 4I)(3, 15, 12) = (-9, -9, 0), and dividing by (1 - 4)(1 + 2) = -9 gives (1, 1, 0). Piece in E_{-2}: subtracting both from \mathbf{w} leaves (1, 2, 2).
- The functions are eigenvectors of D for the distinct eigenvalues 1, 3, -2, so independent by Result 2. g = f' + 2f = 6e^{x} + 5e^{3x}, and then g' - 3g is 6e^{x} + 15e^{3x} minus 18e^{x} + 15e^{3x}, which is -12e^{x}. Dividing by (1 - 3)(1 + 2) = -6 recovers the term 2e^{x}.
- De^{\pm ix} = \pm ie^{\pm ix}, and e^{\pm ix} is never 0, so they are eigenvectors for the distinct eigenvalues i and -i, independent by Result 2. If a\cos x + b\sin x = 0, then with \cos x = \tfrac12(e^{ix} + e^{-ix}) and \sin x = \tfrac{1}{2i}(e^{ix} - e^{-ix}) the coefficients \tfrac{a}{2} + \tfrac{b}{2i} of e^{ix} and \tfrac{a}{2} - \tfrac{b}{2i} of e^{-ix} are 0. Adding gives a = 0, then b = 0.
- (r^k) \in W exactly when r^2 = 3r + 4, so r = 4 or r = -1. These are eigenvectors of the left shift on W for 4 and -1, and \dim W = 2 as in Example 5, so by Result 5 they form a basis. c_1 + c_2 = 2 and 4c_1 - c_2 = 3 give c_1 = c_2 = 1: a_k = 4^k + (-1)^k, with a_2 = 17 and a_3 = 63 = 3 \cdot 17 + 4 \cdot 3.
- Write U_{ij} for the matrix with 1 in row i, column j and 0 elsewhere. Then TU_{11} = TU_{22} = 0, TU_{12} = 3U_{12} and TU_{21} = -3U_{21}. So \dim E_0 \ge 2, \dim E_3 \ge 1, \dim E_{-3} \ge 1, and Result 4 caps the sum at 4: E_0 is the diagonal matrices, E_3 = \operatorname{span}\{U_{12}\}, E_{-3} = \operatorname{span}\{U_{21}\}. Another eigenvalue would push the sum past 4.
- T(x^k) = kx^k - x^k = (k - 1)x^k, so -1, 0, 1, 2 are eigenvalues with eigenvectors 1, x, x^2, x^3. That is 4 distinct eigenvalues on the 4-dimensional space, so by Result 5 there are no others and E_{k-1} = \operatorname{span}\{x^k\}. \ker T = E_0 = \operatorname{span}\{x\}, and \operatorname{im} T is spanned by -1, 0, x^2, 2x^3, so it is \operatorname{span}\{1, x^2, x^3\}.
- \dim E_3 \ge 1 and 2 + 3 + \dim E_3 \le 6, so \dim E_3 = 1. A fourth eigenvalue would make the sum at least 7, so there is none. The sum E_1 \oplus E_2 \oplus E_3 has dimension 6, so it is V (Result 6 of Bases and dimension, proved), and [T]_{\mathcal{B}} = \operatorname{diag}(1, 1, 2, 2, 2, 3).
- No: by Result 4 the dimensions would add to 6 > 5.
- Suppose T(\mathbf{v}_1 + \dots + \mathbf{v}_m) = \mu(\mathbf{v}_1 + \dots + \mathbf{v}_m). Then (\lambda_1 - \mu)\mathbf{v}_1 + \dots + (\lambda_m - \mu)\mathbf{v}_m = \mathbf{0}, and by Result 2 every \lambda_i = \mu. Then \lambda_1 = \lambda_2, against distinctness. (The sum is nonzero by Result 2, so the only way to fail is this one.)
- Let \lambda_1, \dots, \lambda_m be the distinct nonzero eigenvalues, with eigenvectors \mathbf{v}_i. Each \mathbf{v}_i = T(\lambda_i^{-1}\mathbf{v}_i) lies in \operatorname{im} T, and (\mathbf{v}_1, \dots, \mathbf{v}_m) is independent by Result 2. An independent list in \operatorname{im} T has length at most \dim \operatorname{im} T (Result 5 of Bases and dimension, proved), so m \le \operatorname{rank} T. Counting 0 as well gives at most \operatorname{rank} T + 1. A rank-1 operator on \mathbb{F}^5 has at most 2, and \operatorname{diag}(1, 0, 0, 0, 0) has exactly 2.
- If \dim V = n were finite, Result 4 would allow at most n distinct eigenvalues. For the shift L and any \lambda, the sequence \mathbf{s} = (1, \lambda, \lambda^2, \dots) is nonzero and L\mathbf{s} = (\lambda, \lambda^2, \dots) = \lambda\mathbf{s}. When \mathbb{F} is infinite, the first part then shows again that \mathbb{F}^{\infty} is infinite-dimensional.
- Let \mathbf{v}_1, \dots, \mathbf{v}_n be eigenvectors for \lambda_1, \dots, \lambda_n. By Example 4, S\mathbf{v}_k = \mu_k\mathbf{v}_k. Let p = \mu_1p_1 + \dots + \mu_np_n, of degree less than n, with p_k the separating polynomials of \lambda_1, \dots, \lambda_n. Then p(\lambda_k) = \mu_k, so p(T)\mathbf{v}_k = p(\lambda_k)\mathbf{v}_k = \mu_k\mathbf{v}_k = S\mathbf{v}_k (proof of Result 1). Two linear maps that agree on a basis agree everywhere (Result 5 of Linear maps between vector spaces), and (\mathbf{v}_1, \dots, \mathbf{v}_n) is a basis by Result 5, so S = p(T).
- By Result 1, \mathbf{v}_i = p_i(T)\mathbf{v}. Writing p_i(z) = a_0 + a_1z + \dots + a_{m-1}z^{m-1}, which is possible as \deg p_i = m - 1, gives \mathbf{v}_i = a_0\mathbf{v} + a_1T\mathbf{v} + \dots + a_{m-1}T^{m-1}\mathbf{v}.
- U is invariant under T, hence under each T^j (by induction on j), and, being a subspace, under every p(T). So \mathbf{v}_i = p_i(T)\mathbf{v} \in U. If V = E_{\lambda_1} \oplus \dots \oplus E_{\lambda_m}, each \mathbf{u} \in U is \mathbf{u}_1 + \dots + \mathbf{u}_m with \mathbf{u}_i \in E_{\lambda_i}, and each \mathbf{u}_i \in U \cap E_{\lambda_i} by the first part. So U = (U \cap E_{\lambda_1}) + \dots + (U \cap E_{\lambda_m}), and the sum is direct by the zero test, because the sum of the E_{\lambda_i} is.
- T = \operatorname{diag}(1, -1) on \mathbb{R}^2 has the 2 distinct eigenvalues 1 and -1, but T^2 = I has only the eigenvalue 1.