Generalized eigenvectors
When an operator has too few eigenvectors to make a basis, what can take their place?
Where you are 3 in Structure of operators
Why this exists
Diagonalization needs a basis of eigenvectors, and often there is none. The operator T(x, y) = (x + y, y) on \mathbb{F}^2 has the single eigenspace E_1 = \operatorname{span}\{\mathbf{e}_1\}. Yet \mathbf{e}_2 is not lost: (T - I)\mathbf{e}_2 = \mathbf{e}_1, so (T - I)^2\mathbf{e}_2 = \mathbf{0}. Allowing a power of T - \lambda I gives generalized eigenvectors, and over \mathbb{C} they always span V. They cut every complex operator into invariant pieces on which it is \lambda I plus an operator with a zero power, the first half of the Jordan canonical form, and they yield the characteristic polynomial and the Cayley-Hamilton theorem.
Definition
Throughout, V is a finite-dimensional vector space over \mathbb{F} with n = \dim V, T \in \mathcal{L}(V) and \lambda \in \mathbb{F}.
A vector \mathbf{v} is a generalized eigenvector of T for \lambda if \mathbf{v} \neq \mathbf{0} and (T - \lambda I)^k\mathbf{v} = \mathbf{0} for some positive integer k; every eigenvector is one. The generalized eigenspace is
Result 3 shows that G_\lambda consists of \mathbf{0} and every generalized eigenvector for \lambda, however large its k. An operator N is nilpotent if N^k = 0 for some positive integer k. For an eigenvalue \lambda, \dim E_\lambda is its geometric multiplicity and \dim G_\lambda its algebraic multiplicity, or simply multiplicity.
The determinant of T is \det T = \det[T]_{\mathcal{B}} for any basis \mathcal{B}: another basis gives P^{-1}[T]_{\mathcal{B}}P (Result 5 of The matrix of a linear map), with the same determinant by the product rule (Properties of determinants: row operations, products, transposes). The characteristic polynomial is p_T(\lambda) = \det(T - \lambda I), as in The characteristic equation and finding eigenvalues; by Result 7, over \mathbb{C} the multiplicity of \lambda as its zero is \dim G_\lambda.
Intuition
Fix \lambda and apply S = T - \lambda I again and again. The null spaces climb,
and in dimension n they can grow at most n times. Once two neighbors agree the chain has stopped for good (Result 1), and where it stops is G_\lambda. A vector of \ker S^2 outside \ker S is sent by S to an eigenvector; a vector of \ker S^3 outside \ker S^2 is sent into \ker S^2; and so on. So on G_\lambda, T is \lambda I plus S, and S dies after at most \dim G_\lambda steps. Over \mathbb{C} these pieces fill V with no overlap (Result 5), and a basis of each makes [T] block diagonal with one triangular block per eigenvalue.
Results
V is finite-dimensional, n = \dim V and T \in \mathcal{L}(V). Results 1 to 4 hold over any field with the same proofs. Results 5 and 6 need m_T to be a product of linear factors, always true over \mathbb{C} by the The fundamental theorem of algebra.
Result 1 (null spaces of powers stop growing). (a) \ker T^k \subseteq \ker T^{k+1} for every k \ge 0. (b) If \ker T^m = \ker T^{m+1}, then \ker T^m = \ker T^{m+k} for every k \ge 0. (c) \ker T^n = \ker T^{n+k} for every k \ge 0.
Result 2 (a power splits the space). V = \ker T^n \oplus \operatorname{im} T^n.
Result 3 (generalized eigenspaces). (a) G_\lambda is the set of \mathbf{v} with (T - \lambda I)^k\mathbf{v} = \mathbf{0} for some k \ge 1. (b) G_\lambda is invariant under T and contains E_\lambda. (c) G_\lambda \neq \{\mathbf{0}\} if and only if \lambda is an eigenvalue. (d) With d = \dim G_\lambda, (T - \lambda I)^d is zero on G_\lambda.
Result 4 (independence). If \lambda_1, \dots, \lambda_m are distinct and \mathbf{v}_j is a generalized eigenvector for \lambda_j, then (\mathbf{v}_1, \dots, \mathbf{v}_m) is linearly independent, and G_{\lambda_1} + \dots + G_{\lambda_m} is direct.
Result 5 (generalized eigenspace decomposition). Let V \neq \{\mathbf{0}\}, let m_T be a product of linear factors, and let \lambda_1, \dots, \lambda_m be the distinct eigenvalues of T. (a) V = G_{\lambda_1} \oplus \dots \oplus G_{\lambda_m}, so V has a basis of generalized eigenvectors. (b) Some basis \mathcal{B} gives
(zeros off the blocks), with A_j upper triangular of size \dim G_{\lambda_j} and every diagonal entry \lambda_j.
Result 6 (multiplicities and the minimal polynomial). Under the hypotheses of Result 5, let d_j = \dim G_{\lambda_j} and r_j the least k with \ker(T - \lambda_jI)^k = G_{\lambda_j}. Then (a) d_1 + \dots + d_m = n; (b) 1 \le r_j \le d_j and
Result 7 (characteristic polynomial and Cayley-Hamilton). Let V \neq \{\mathbf{0}\} be complex, notation as in Result 6. (a) The characteristic polynomial is
so \lambda_j is a zero of p_T of multiplicity \dim G_{\lambda_j}. (b) (Cayley-Hamilton) p_T(T) = 0 and m_T \mid p_T. (c) Every real or complex n \times n matrix A satisfies p_A(A) = 0, where p_A(\lambda) = \det(A - \lambda I_n).
Why it is true
Result 1. (a) If T^k\mathbf{v} = \mathbf{0}, then T^{k+1}\mathbf{v} = T(T^k\mathbf{v}) = \mathbf{0}. (b) If T^{m+k+1}\mathbf{v} = \mathbf{0}, then T^k\mathbf{v} \in \ker T^{m+1} = \ker T^m, so T^{m+k}\mathbf{v} = \mathbf{0}. With (a), \ker T^{m+k+1} = \ker T^{m+k} for every k, and these equalities chain. (c) If \ker T^n \neq \ker T^{n+1}, then by (b) no two neighbors among \ker T^0, \dots, \ker T^{n+1} are equal, so by (a) each is a proper subspace of the next, of smaller dimension (Result 6 of Bases and dimension, proved). The dimensions rise from 0 at each of n + 1 steps, so \dim\ker T^{n+1} > n, which is impossible. So \ker T^n = \ker T^{n+1}, and (b) finishes.
Result 2. Let \mathbf{v} = T^n\mathbf{u} lie in \ker T^n. Then T^{2n}\mathbf{u} = \mathbf{0}, so \mathbf{u} \in \ker T^{2n} = \ker T^n by Result 1(c), and \mathbf{v} = \mathbf{0}. The sum is direct (Result 3 of Sums and direct sums of subspaces), and by Result 6 there and the fundamental theorem of linear maps (Null space, range and the fundamental theorem, Result 5) it has dimension \dim\ker T^n + \dim\operatorname{im} T^n = n, so it is V.
Result 3. Write S = T - \lambda I. (a) By Result 1(a) and (c) for S, \ker S^k \subseteq \ker S^n for every k. (b) S^n commutes with T (Result 5 of Invariant subspaces), so S^n\mathbf{v} = \mathbf{0} gives S^nT\mathbf{v} = TS^n\mathbf{v} = \mathbf{0}; and E_\lambda = \ker S \subseteq \ker S^n. (c) If \lambda is an eigenvalue, \{\mathbf{0}\} \neq E_\lambda \subseteq G_\lambda. Conversely, if \mathbf{v} \in G_\lambda is nonzero and k \ge 1 is least with S^k\mathbf{v} = \mathbf{0}, then S^{k-1}\mathbf{v} is an eigenvector for \lambda. (d) By (b), N = S|_{G_\lambda} is an operator on G_\lambda with N^n = 0, and Result 1(c) for N on the d-dimensional G_\lambda gives \ker N^d = \ker N^n = G_\lambda.
Result 4. Let a_1\mathbf{v}_1 + \dots + a_m\mathbf{v}_m = \mathbf{0}; we show a_1 = 0, and each a_j is alike. Let k \ge 0 be largest with \mathbf{w} = (T - \lambda_1I)^k\mathbf{v}_1 \neq \mathbf{0} (it exists by Result 3(a)). Then T\mathbf{w} = \lambda_1\mathbf{w}, so (T - \lambda_jI)\mathbf{w} = (\lambda_1 - \lambda_j)\mathbf{w}. Apply
whose factors commute. Moving (T - \lambda_jI)^n to the right shows R\mathbf{v}_j = \mathbf{0} for j \ge 2, while R\mathbf{v}_1 is the nonzero vector (\lambda_1 - \lambda_2)^n\cdots(\lambda_1 - \lambda_m)^n\mathbf{w}. So a_1 = 0. If \mathbf{0} = \mathbf{u}_1 + \dots + \mathbf{u}_m with \mathbf{u}_j \in G_{\lambda_j}, the nonzero \mathbf{u}_j would satisfy a relation with coefficients 1, so there are none, and the sum is direct (Result 2 of Sums and direct sums of subspaces).
Result 5. (a) Induction on n, over operators whose minimal polynomial is a product of linear factors (for V = \{\mathbf{0}\} read “the empty sum is V”). Let n \ge 1. A zero \lambda_1 of m_T is an eigenvalue (Result 3 of The minimal polynomial). Result 2 for T - \lambda_1I gives V = G_{\lambda_1} \oplus U with U = \operatorname{im}(T - \lambda_1I)^n, invariant since T(T - \lambda_1I)^n\mathbf{x} = (T - \lambda_1I)^nT\mathbf{x}, and \dim U < n. As m_T(T|_U) = 0, Result 2(b) of The minimal polynomial makes m_{T|_U} a product of linear factors, so by induction U is the sum of the generalized eigenspaces of T|_U. Each eigenvalue \mu of T|_U is one of T (Result 4 of Invariant subspaces), and (T|_U - \mu I)^k\mathbf{u} = \mathbf{0} puts \mathbf{u} in G_\mu by Result 3(a). So V = G_{\lambda_1} + U \subseteq G_{\lambda_1} + \dots + G_{\lambda_m}, a direct sum by Result 4, and joined bases of the summands form a basis (Result 5 of Sums and direct sums of subspaces).
(b) By Result 3(d), (z - \lambda_j)^{d_j} annihilates T|_{G_{\lambda_j}}, so by Results 2(b) and 6(b) of The minimal polynomial some basis of G_{\lambda_j} makes its matrix A_j upper triangular. The diagonal entries are eigenvalues of T|_{G_{\lambda_j}} (Upper-triangular matrices over the complex numbers, Result 4), so zeros of its minimal polynomial, a power of z - \lambda_j; all equal \lambda_j. Join the bases; as each G_{\lambda_j} is invariant, the columns for its basis vectors are zero outside block j.
Result 6. (a) The joined basis has d_1 + \dots + d_m vectors. (b) \ker(T - \lambda_jI)^k \subseteq G_{\lambda_j} for every k (Result 3(a)), with equality at k = d_j (Result 3(d)) but not at k = 0; so 1 \le r_j \le d_j, and by Result 1(a) (T - \lambda_jI)^k vanishes on G_{\lambda_j} exactly when k \ge r_j. Let q(z) = (z - \lambda_1)^{r_1}\cdots(z - \lambda_m)^{r_m}. On each G_{\lambda_j}, moving (T - \lambda_jI)^{r_j} to the right shows q(T) = 0 there, so q(T) = 0 by Result 5(a), and m_T = (z - \lambda_1)^{s_1}\cdots(z - \lambda_m)^{s_m} with every s_j \le r_j (Result 2(b) of The minimal polynomial). Suppose s_k < r_k, and pick \mathbf{v} \in G_{\lambda_k} with \mathbf{w} = (T - \lambda_kI)^{s_k}\mathbf{v} \neq \mathbf{0}. For j \neq k, T - \lambda_jI maps G_{\lambda_k} into itself injectively, since its null space there lies in G_{\lambda_j} \cap G_{\lambda_k} = \{\mathbf{0}\}. So m_T(T)\mathbf{v} = \prod_{j \neq k}(T - \lambda_jI)^{s_j}\mathbf{w} \neq \mathbf{0}, which is impossible. Hence s_j = r_j.
Result 7. (a) In the basis of Result 5(b), [T]_{\mathcal{B}} - \lambda I_n is upper triangular with diagonal entries \lambda_j - \lambda, each d_j times, and a triangular determinant is the product of the diagonal (Determinants by cofactor expansion). With d_1 + \dots + d_m = n,
The other factors do not vanish at \lambda_j, so its multiplicity is exactly d_j. (b) By Result 6(b), p_T = (-1)^nm_Ts with s(z) = \prod_j(z - \lambda_j)^{d_j - r_j}, so m_T \mid p_T and p_T(T) = (-1)^ns(T)m_T(T) = 0. (c) A complex A is the matrix of \mathbf{x} \mapsto A\mathbf{x} on \mathbb{C}^n, and p(T) = 0 exactly when p([T]_{\mathcal{B}}) = 0 (Example 5 of The minimal polynomial), so (b) gives p_A(A) = 0. A real matrix is also complex with the same p_A, so the equation holds for it, and so for operators on real spaces through their matrices.
Worked examples
Climbing to a generalized eigenspace
Let A = \begin{bmatrix} 2 & 1 & 0 \\ -1 & 4 & 0 \\ -2 & 2 & 1 \end{bmatrix}. Find E_\lambda and G_\lambda for each eigenvalue, and a basis of \mathbb{F}^3 of generalized eigenvectors.
SolutionExpanding along the third column, p_A(\lambda) = (1 - \lambda)\big[(2 - \lambda)(4 - \lambda) + 1\big], which is (1 - \lambda)(\lambda - 3)^2, so the eigenvalues are 3 and 1. Next,
The rows of the first force x_2 = x_1, x_3 = 0, so E_3 = \operatorname{span}\{(1, 1, 0)\}. The second has rank 1, so \ker(A - 3I_3)^2 is the plane x_1 - x_2 + x_3 = 0, with basis (1, 1, 0), (-1, 0, 1). No higher power is needed: G_1 \neq \{\mathbf{0}\} and G_3 \cap G_1 = \{\mathbf{0}\} (Results 3(c) and 4), so \dim G_3 \le 2.
For 1, the first two rows of A - I_3 = \begin{bmatrix} 1 & 1 & 0 \\ -1 & 3 & 0 \\ -2 & 2 & 0 \end{bmatrix} force x_1 = x_2 = 0, so E_1 = \operatorname{span}\{(0, 0, 1)\}, and \dim G_1 \le 1 gives G_1 = E_1. So \mathbb{F}^3 = G_3 \oplus G_1 with basis \big((1, 1, 0), (-1, 0, 1), (0, 0, 1)\big). The middle vector is one step from an eigenvector, (A - 3I_3)(-1, 0, 1) = (1, 1, 0), and the matrix in this basis is \begin{bmatrix} 3 & 1 & 0 \\ 0 & 3 & 0 \\ 0 & 0 & 1 \end{bmatrix}, as Result 5(b) promises.
Differentiation on exponentials times polynomials
Let V = \operatorname{span}\{e^{2x}, xe^{2x}, x^2e^{2x}\}, three-dimensional, and Df = f'. Find the null spaces of the powers of D - 2I, then G_2, m_D and p_D.
SolutionBy the product rule, (D - 2I)(x^ke^{2x}) = kx^{k-1}e^{2x}, so V is invariant and in the basis (e^{2x}, xe^{2x}, x^2e^{2x})
triangular with only 2 on the diagonal, the only eigenvalue (Upper-triangular matrices over the complex numbers, Result 4). For f = (a + bx + cx^2)e^{2x},
So the null spaces are \operatorname{span}\{e^{2x}\}, \operatorname{span}\{e^{2x}, xe^{2x}\} and V: the chain stops only at k = 3 = \dim V, the latest Result 1 allows. Thus G_2 = V, E_2 is a line, r = 3, and by Results 6 and 7, m_D(z) = (z - 2)^3 and p_D(\lambda) = -(\lambda - 2)^3.
Amara’s fund at the bookbindery
Amara is saving for a board shear for her bookbindery. The fund earns 5\% a year, and at the end of each year she adds a deposit 5\% larger than the one before. Let b_k be the balance at the start of year k and d_k the deposit at its end:
The fund starts empty and d_0 = \$2{,}000. Find b_k, and b_{10} to the cent.
SolutionWith \mathbf{x}_k = (b_k, d_k) and r = \tfrac{21}{20}, \mathbf{x}_k = A^k\mathbf{x}_0 for A = \begin{bmatrix} r & 1 \\ 0 & r \end{bmatrix}. Let N = A - rI_2 = \begin{bmatrix} 0 & 1 \\ 0 & 0 \end{bmatrix}. As N^2 = 0, G_r = \mathbb{R}^2 while E_r = \operatorname{span}\{\mathbf{e}_1\}, so Diagonalization cannot produce A^k. The generalized eigenvector \mathbf{e}_2 can: rI_2 and N commute, so the binomial theorem expands (rI_2 + N)^k, and every term containing N^2 vanishes:
With \mathbf{x}_0 = (0, 2000), b_k = 2000k\big(\tfrac{21}{20}\big)^{k-1}. Check: b_2 = \tfrac{21}{20}\cdot 2000 + 2100 = 4200. Then b_{10} = 20{,}000(1.05)^9 \approx \$31{,}026.56.
One eigenvalue means nilpotent after a shift
Let V \neq \{\mathbf{0}\} be complex and \lambda \in \mathbb{C}. Prove that \lambda is the only eigenvalue of T if and only if T - \lambda I is nilpotent, and that then (T - \lambda I)^{n} = 0.
SolutionIf \lambda is the only eigenvalue, Result 5(a) with m = 1 gives V = G_\lambda = \ker(T - \lambda I)^n, so (T - \lambda I)^n = 0. Conversely, let (T - \lambda I)^k = 0 with k \ge 1. If T\mathbf{v} = \mu\mathbf{v} with \mathbf{v} \neq \mathbf{0}, then \mathbf{0} = (T - \lambda I)^k\mathbf{v} = (\mu - \lambda)^k\mathbf{v}, so \mu = \lambda. As T has an eigenvalue (Result 1 of Upper-triangular matrices over the complex numbers), \lambda is its only one.
Multiplicities and Cayley-Hamilton in dimension four
Let A = \begin{bmatrix} 1 & 1 & 0 & 0 \\ -1 & 3 & 0 & 0 \\ 0 & 0 & 2 & 0 \\ 3 & -3 & 3 & -1 \end{bmatrix}. Find p_A, both multiplicities of each eigenvalue, and m_A, and confirm p_A(A) = 0.
SolutionThe fourth column of A - \lambda I_4 has only the entry -1 - \lambda, and the remaining block has a lone 2 - \lambda in its third column, so
By Result 7(a), \dim G_2 = 3 and \dim G_{-1} = 1. The rows of A - 2I_4 are (-1, 1, 0, 0) twice, \mathbf{0} and (3, -3, 3, -3), forcing x_2 = x_1 and x_4 = x_3: E_2 has basis (1, 1, 0, 0), (0, 0, 1, 1). The only nonzero row of (A - 2I_4)^2 is row 4, (-9, 9, -9, 9), so \ker(A - 2I_4)^2 is the three-dimensional x_1 - x_2 + x_3 - x_4 = 0, all of G_2, adding (1, 0, 0, 1) to the basis. So 2 has geometric multiplicity 2, algebraic multiplicity 3, and r = 2. Column 4 of A + I_4 is zero, so G_{-1} = E_{-1} = \operatorname{span}\{\mathbf{e}_4\}. By Result 6(b),
Directly, (A - 2I_4)^2(A + I_4) can be nonzero only in row 4: (-9, 9, -9, 9) times the columns (2, -1, 0, 3), (1, 4, 0, -3), (0, 0, 3, 3), \mathbf{0} of A + I_4 gives -18 - 9 + 27 = 0, -9 + 36 - 27 = 0, -27 + 27 = 0 and 0. Then p_A(A) = (A - 2I_4)\,m_A(A) = 0, as Cayley-Hamilton requires.
Boundaries
- Over \mathbb{R} the pieces may not fill V. The quarter turn (x, y) \mapsto (-y, x) has minimal polynomial z^2 + 1 and no real eigenvalue, so every G_\lambda is \{\mathbf{0}\}. Cayley-Hamilton still holds over \mathbb{R} (Result 7(c)).
- Finite dimension is needed. On \mathcal{P}(\mathbb{F}), \ker D^k = \mathcal{P}_{k-1}(\mathbb{F}) for k \ge 1 never stops growing; every nonzero polynomial is a generalized eigenvector for 0, yet no power of D is zero.
- A smaller power may not split V. For T(x, y) = (y, 0), \ker T = \operatorname{im} T = \operatorname{span}\{\mathbf{e}_1\}, so \ker T + \operatorname{im} T is not direct; Result 2 needs the power n.
- Multiplicities and m_T do not determine T. Write J_k(\lambda) for the k \times k matrix with \lambda on the diagonal, 1 on the superdiagonal and 0 elsewhere. \operatorname{diag}(J_2(0), J_2(0)) and \operatorname{diag}(J_2(0), 0, 0) share p(\lambda) = \lambda^4 and m(z) = z^2, but \dim E_0 is 2 and 3, so they are not similar. Jordan canonical form settles this.
Where this goes next
Jordan canonical form chooses inside each G_\lambda a basis of chains (T - \lambda I)^{k-1}\mathbf{v}, \dots, \mathbf{v}, turning each block of Result 5(b) into Jordan blocks J_k(\lambda), with r_j the size of the largest. Trace and determinant from the structure uses Result 7(a) to make the trace and determinant the sum and product of the eigenvalues, each counted \dim G_\lambda times. When an operator is diagonalizable is the case where every r_j = 1 (Exercise 11).
Exercises
Generalized eigenspaces of matrices
- Let A = \begin{bmatrix} 5 & -1 \\ 1 & 3 \end{bmatrix}. Find p_A, E_4, G_4, and a basis (\mathbf{v}_1, \mathbf{v}_2) of \mathbb{F}^2 with \mathbf{v}_1 an eigenvector and (A - 4I_2)\mathbf{v}_2 = \mathbf{v}_1.
- Let C = \begin{bmatrix} 1 & -1 & 0 \\ 1 & -1 & 0 \\ 0 & 0 & 3 \end{bmatrix}. Find \ker C, \operatorname{im} C, \ker C^3 and \operatorname{im} C^3. Which of \ker C + \operatorname{im} C and \ker C^3 + \operatorname{im} C^3 is direct?
- Let M = \begin{bmatrix} 1 & 1 & 0 & 2 \\ 0 & 1 & 1 & 0 \\ 0 & 0 & 1 & 3 \\ 0 & 0 & 0 & 4 \end{bmatrix}. Find \dim\ker(M - I_4)^k for k = 1, 2, 3, a basis of each generalized eigenspace, p_M and m_M.
- Let R = \begin{bmatrix} 0 & -1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 2 \end{bmatrix}. Over \mathbb{R}, find every nonzero generalized eigenspace and decide whether they add up to \mathbb{R}^3. Then do the same over \mathbb{C}.
Operators and applications
- Let V = \operatorname{span}\{\cos x, \sin x, x\cos x, x\sin x\}, four-dimensional over \mathbb{C}, and Df = f'. Find the eigenvalues of D and, for each, a basis of E_\lambda and of G_\lambda. (Use e^{ix} = \cos x + i\sin x.)
- Let A = \begin{bmatrix} 1 & 4 \\ -1 & 5 \end{bmatrix}. Show that A - 3I_2 is nilpotent and use it to compute A^5.
- Amara’s second fund, for a cutting table, earns 4\% a year, each deposit is 4\% larger than the last, the first deposit is \$1{,}500, and the fund starts with \$3{,}000. With the model of Example 3, find the balance at the start of year 5, to the cent.
- An operator T on \mathbb{C}^7 has only the eigenvalues 2 and -3, and \dim\ker(T - 2I)^k is 2, 4, 5, 5 for k = 1, 2, 3, 4. Find \dim G_2, \dim G_{-3}, p_T and the exponent of z - 2 in m_T. Do the data decide the exponent of z + 3?
- An operator T on \mathbb{C}^3 has \dim G_1 = 2 and \dim G_{-2} = 1. Find p_T, and use Cayley-Hamilton to write T^{-1} as a polynomial in T.
Proofs
- Let S, T \in \mathcal{L}(V) commute. Prove that S maps each generalized eigenspace G_\lambda of T into itself.
- Let V be complex. Prove that T has a basis of eigenvectors if and only if \ker(T - \lambda I)^2 = \ker(T - \lambda I) for every \lambda \in \mathbb{C}.
- Let \lambda be an eigenvalue of T. Prove that \lambda is the only eigenvalue of the restriction T|_{G_\lambda}.
- Let c_k = \dim\ker T^k. Prove that c_{k+1} - c_k \le c_k - c_{k-1} for every k \ge 1.
- Let V be complex and p a polynomial. Prove that p(T) is nilpotent if and only if p(\lambda) = 0 for every eigenvalue \lambda of T.
Answers
- p_A(\lambda) = (5 - \lambda)(3 - \lambda) + 1 = (\lambda - 4)^2. A - 4I_2 = \begin{bmatrix} 1 & -1 \\ 1 & -1 \end{bmatrix} has rank 1 and square 0, so E_4 = \operatorname{span}\{(1, 1)\} and G_4 = \mathbb{F}^2. Take \mathbf{v}_1 = (1, 1), \mathbf{v}_2 = (1, 0): (A - 4I_2)(1, 0) = (1, 1).
- C\mathbf{x} = (x_1 - x_2, x_1 - x_2, 3x_3), so \ker C = \operatorname{span}\{(1, 1, 0)\} and \operatorname{im} C = \operatorname{span}\{(1, 1, 0), \mathbf{e}_3\}; they share (1, 1, 0), so that sum is not direct. The upper-left block squares to 0, so C^3 = \operatorname{diag}(0, 0, 27): \ker C^3 = \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_2\} and \operatorname{im} C^3 = \operatorname{span}\{\mathbf{e}_3\}, a direct sum equal to \mathbb{F}^3.
- Dimensions 1, 2, 3: M - I_4 sends \mathbf{e}_3 \mapsto \mathbf{e}_2 \mapsto \mathbf{e}_1 \mapsto \mathbf{0}, and the fourth coordinate of (M - I_4)^k\mathbf{x} is 3^kx_4, so the null space is \operatorname{span}\{\mathbf{e}_1, \dots, \mathbf{e}_k\}. M is triangular, so p_M(\lambda) = (\lambda - 1)^3(\lambda - 4), G_1 = \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_2, \mathbf{e}_3\} and G_4 = \operatorname{span}\{(7, 3, 9, 9)\} (Result 7). As r_1 = 3, m_M(z) = (z - 1)^3(z - 4).
- p_R(\lambda) = (2 - \lambda)(\lambda^2 + 1). Over \mathbb{R} only 2 is an eigenvalue; on \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_2\}, R - 2I_3 acts by the invertible \begin{bmatrix} -2 & -1 \\ 1 & -2 \end{bmatrix}, so G_2 = \operatorname{span}\{\mathbf{e}_3\}, a line, not \mathbb{R}^3. Over \mathbb{C} each of 2, i, -i has \dim G = 1: G_i = \operatorname{span}\{(1, -i, 0)\}, G_{-i} = \operatorname{span}\{(1, i, 0)\}, and \mathbb{C}^3 = G_2 \oplus G_i \oplus G_{-i}.
- Eigenvalues \pm i. (D - iI)(ae^{ix} + bxe^{ix}) = be^{ix}, so \operatorname{span}\{e^{ix}, xe^{ix}\} \subseteq G_i; likewise \operatorname{span}\{e^{-ix}, xe^{-ix}\} \subseteq G_{-i}. These functions lie in V, and the two planes already fill its four dimensions (Result 4), so they are G_i and G_{-i}, and there is no other eigenvalue. As E_i \subseteq G_i, the formula (b = 0) gives E_i = \operatorname{span}\{e^{ix}\}, and likewise E_{-i} = \operatorname{span}\{e^{-ix}\}.
- N = A - 3I_2 = \begin{bmatrix} -2 & 4 \\ -1 & 2 \end{bmatrix}, N^2 = 0, so A^5 = 243I_2 + 405N = \begin{bmatrix} -567 & 1620 \\ -405 & 1053 \end{bmatrix}.
- With r = 1.04, Example 3 gives b_k = r^kb_0 + kr^{k-1}d_0, so b_5 = 3000(1.04)^5 + 7500(1.04)^4, about 3649.96 + 8773.94 = \$12{,}423.90.
- The chain stops at k = 3 (Result 1(b)), so \dim G_2 = 5 and the exponent of z - 2 is 3 (Result 6). \dim G_{-3} = 7 - 5 = 2, and p_T(\lambda) = -(\lambda - 2)^5(\lambda + 3)^2. The exponent of z + 3 is not decided: \operatorname{diag}(J_3(2), J_2(2), -3, -3) and \operatorname{diag}(J_3(2), J_2(2), J_2(-3)) both fit, with exponents 1 and 2.
- p_T(\lambda) = -(\lambda - 1)^2(\lambda + 2), which is -(\lambda^3 - 3\lambda + 2). So T^3 - 3T + 2I = 0, T(3I - T^2) = 2I and T^{-1} = \tfrac{1}{2}(3I - T^2).
- S commutes with T - \lambda I, as ST - \lambda S = TS - \lambda S, and so with each power: S(T - \lambda I)^{k+1} = (T - \lambda I)^kS(T - \lambda I) = (T - \lambda I)^{k+1}S by induction. If \mathbf{v} \in G_\lambda, then (T - \lambda I)^nS\mathbf{v} = S(T - \lambda I)^n\mathbf{v} = \mathbf{0}, so S\mathbf{v} \in G_\lambda.
- Write S_\lambda = T - \lambda I; if V = \{\mathbf{0}\}, both conditions hold. If \ker S_\lambda^2 = \ker S_\lambda for every \lambda, Result 1(b) with m = 1 gives G_\lambda = E_\lambda, so by Result 5(a) V = E_{\lambda_1} \oplus \dots \oplus E_{\lambda_m}, and joined bases of the eigenspaces are a basis of eigenvectors. Conversely, a basis of eigenvectors gives V = E_{\lambda_1} + \dots + E_{\lambda_m}, so n \le \sum\dim E_{\lambda_j} \le \sum\dim G_{\lambda_j} = n (Result 6(a)). Hence each E_{\lambda_j} = G_{\lambda_j}, which contains \ker S_{\lambda_j}^2. For \lambda not an eigenvalue, S_\lambda and S_\lambda^2 are injective.
- G_\lambda is invariant (Result 3(b)) and contains an eigenvector for \lambda, so \lambda is an eigenvalue of T|_{G_\lambda}. If T\mathbf{v} = \mu\mathbf{v} with \mathbf{0} \neq \mathbf{v} \in G_\lambda, then \mathbf{0} = (T - \lambda I)^n\mathbf{v} = (\mu - \lambda)^n\mathbf{v}, so \mu = \lambda.
- By Result 7 of Sums and direct sums of subspaces, \ker T^{k+1} = \ker T^k \oplus W with \dim W = c_{k+1} - c_k. For \mathbf{w} \in W, T^k(T\mathbf{w}) = \mathbf{0}, so T(W) \subseteq \ker T^k. If T\mathbf{w} \in \ker T^{k-1}, then T^k\mathbf{w} = \mathbf{0}, so \mathbf{w} \in W \cap \ker T^k = \{\mathbf{0}\}. Hence T|_W is injective and T(W) \cap \ker T^{k-1} = \{\mathbf{0}\}, so the direct sum T(W) \oplus \ker T^{k-1} lies in \ker T^k and (c_{k+1} - c_k) + c_{k-1} \le c_k.
- If p(T)^k = 0 and T\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} \neq \mathbf{0}, then \mathbf{0} = p(T)^k\mathbf{v} = p(\lambda)^k\mathbf{v}, so p(\lambda) = 0. Conversely, let p vanish at every eigenvalue \lambda_j. By the The remainder and factor theorems, p(z) = (z - \lambda_j)s_j(z), so p(T)^n = s_j(T)^n(T - \lambda_jI)^n, which is zero on G_{\lambda_j}. By Result 5(a), p(T)^n = 0 (if V = \{\mathbf{0}\} there is nothing to prove).