The minimal polynomial
The powers of an operator on a finite-dimensional space cannot all be independent. What does the first dependence among them say about the operator?
Where you are 5 in Eigenvalues, proved
Why this exists
Let V be finite-dimensional with \dim V = n and let T \in \mathcal{L}(V). The operators I, T, T^2, \dots all live in \mathcal{L}(V), a space of dimension n^2 (Invertibility and isomorphic spaces), so the list cannot stay independent forever: some nonzero polynomial p has p(T) = 0. The first such dependence, the minimal polynomial, has the eigenvalues as its zeros, writes T^{-1} and every power of T as a short polynomial in T, and decides over any field whether T has an upper-triangular matrix, which Upper-triangular matrices over the complex numbers settled only over \mathbb{C}.
Definition
Throughout, V is a finite-dimensional vector space over \mathbb{F} and T \in \mathcal{L}(V). A polynomial is monic if its leading coefficient is 1; the zero polynomial is not monic. For p(z) = a_0 + a_1z + \dots + a_mz^m, p(T) = a_0I + a_1T + \dots + a_mT^m as in Invariant subspaces, and p annihilates T when p(T) = 0. A polynomial q divides p, written q \mid p, when p = qs for some polynomial s.
The minimal polynomial m_T is the monic polynomial of least degree that annihilates T; Result 1 shows there is exactly one. For a square matrix A \in \mathbb{F}^{n \times n}, m_A is the minimal polynomial of the operator \mathbf{x} \mapsto A\mathbf{x} on \mathbb{F}^n.
On the zero space I = 0, so the constant polynomial 1 annihilates T and m_T = 1 (some books leave this space out). On a nonzero space 1(T) = I \neq 0, so \deg m_T \ge 1.
Intuition
Walk along the list I, T, T^2, \dots in \mathcal{L}(V) and stop at the first power that is a combination of the earlier ones:
Then m_T(z) = z^d + c_{d-1}z^{d-1} + \dots + c_0: a monic annihilator of smaller degree would have made an earlier power a combination of the ones before it (independence of a list, here of operators).
Eigenvalues appear at once. If T\mathbf{v} = \lambda\mathbf{v}, then p(T)\mathbf{v} = p(\lambda)\mathbf{v} for every polynomial p (the first step in the proof of Result 1 of Eigenvectors for distinct eigenvalues are independent). With p = m_T and \mathbf{v} \neq \mathbf{0} this gives m_T(\lambda) = 0. The converse uses that m_T is the least annihilator.
Results
V is finite-dimensional and T \in \mathcal{L}(V). Every result holds over any field with the same proof, except Result 3(b), which needs \mathbb{F} = \mathbb{C}.
Result 1 (existence, uniqueness, degree). There is exactly one monic polynomial of least degree that annihilates T, and its degree is at most \dim V.
Result 2 (annihilators are multiples). Let q \in \mathcal{P}(\mathbb{F}). (a) q(T) = 0 if and only if m_T \mid q. (b) If q(z) = (z - \lambda_1)\cdots(z - \lambda_k) and q(T) = 0, then m_T is the product of some of these factors.
Result 3 (zeros are eigenvalues). (a) A scalar \lambda \in \mathbb{F} is an eigenvalue of T if and only if m_T(\lambda) = 0. (b) If V \neq \{\mathbf{0}\} is complex, T has an eigenvalue.
Result 4 (invertibility). T is invertible if and only if m_T(0) \neq 0. In that case, writing m_T(z) = z^d + c_{d-1}z^{d-1} + \dots + c_1z + c_0,
Result 5 (restrictions). If U is a subspace invariant under T, then m_{T|_U} \mid m_T.
Result 6 (the triangular test). (a) If [T]_{\mathcal{B}} is upper triangular with diagonal \lambda_1, \dots, \lambda_n, then (T - \lambda_1I)\cdots(T - \lambda_nI) = 0. (b) T has an upper-triangular matrix with respect to some basis of V if and only if m_T(z) = (z - \mu_1)\cdots(z - \mu_r) for some \mu_1, \dots, \mu_r \in \mathbb{F}, repeats allowed.
Why it is true
Two matrix facts. (i) For any basis \mathcal{B} of V, p(T) = 0 exactly when p([T]_{\mathcal{B}}) = 0 (Example 5, whose proof uses none of the results here). (ii) If A, A' are k \times k and Y, Y' are square of one size, block multiplication gives
By induction on powers, then adding, p\Big(\begin{bmatrix} A & X \\ 0 & Y \end{bmatrix}\Big) has diagonal blocks p(A), p(Y) and a zero lower-left block.
Result 1. First, a monic annihilator of degree at most n = \dim V exists, by Strong induction and well-ordering on n. If n = 0, the polynomial 1 works. Let n \ge 1 and assume the claim in every smaller dimension. Pick \mathbf{u} \neq \mathbf{0}. The list (\mathbf{u}, T\mathbf{u}, \dots, T^n\mathbf{u}) has n + 1 entries, so it is dependent (Bases and dimension, proved), and by the linear dependence lemma (Span and independence, proved) some entry lies in the span of those before it; let T^k\mathbf{u} be the first such entry. Then k \ge 1 because \mathbf{u} \neq \mathbf{0}, and (\mathbf{u}, \dots, T^{k-1}\mathbf{u}) is independent, since each of its entries lies outside the span of those before it (Result 3 there). Write T^k\mathbf{u} = -(c_0\mathbf{u} + \dots + c_{k-1}T^{k-1}\mathbf{u}) and q(z) = z^k + c_{k-1}z^{k-1} + \dots + c_0, so q(T)\mathbf{u} = \mathbf{0}.
T maps Z = \operatorname{span}\{\mathbf{u}, \dots, T^{k-1}\mathbf{u}\} into itself: it sends T^j\mathbf{u} to T^{j+1}\mathbf{u}, which lies in Z for j < k - 1, and for j = k - 1 by the choice of k. Extend (\mathbf{u}, \dots, T^{k-1}\mathbf{u}) to a basis \mathcal{B} of V. By Result 4 of Invariant subspaces, [T]_{\mathcal{B}} = \begin{bmatrix} A & X \\ 0 & Y \end{bmatrix} with A of size k \times k and Y of size (n - k) \times (n - k). Polynomials in T commute (Result 5 of Invariant subspaces), so q(T)T^j\mathbf{u} = T^jq(T)\mathbf{u} = \mathbf{0} for every j, and the first k columns of [q(T)]_{\mathcal{B}} = q([T]_{\mathcal{B}}) are zero. By fact (ii) those columns hold q(A) above a zero block, so q(A) = 0. As n - k < n, the induction hypothesis for \mathbf{x} \mapsto Y\mathbf{x} on \mathbb{F}^{n-k}, with fact (i), gives a monic s with \deg s \le n - k and s(Y) = 0 (if k = n, the block Y is empty and s = 1). By fact (ii),
This product is the matrix of q(T)s(T) = (qs)(T), so qs is monic, annihilates T and has degree at most n.
Since monic annihilators exist, one of least degree d exists. If m and \hat{m} are two such, then m - \hat{m} annihilates T and, as the leading terms cancel, is zero or of degree below d. If it were nonzero, dividing by its leading coefficient would give a monic annihilator of degree below d. So m = \hat{m}.
Result 2. (a) If q = m_Ts, then q(T) = s(T)m_T(T) = 0 by Result 5 of Invariant subspaces. Conversely, let q(T) = 0 and divide (Polynomial long division): q = m_Ts + r with r = 0 or \deg r < \deg m_T. Then r(T) = q(T) - s(T)m_T(T) = 0, and a nonzero r divided by its leading coefficient would be a monic annihilator of smaller degree than m_T. So r = 0.
(b) By (a), m_Ts = (z - \lambda_1)\cdots(z - \lambda_k) for some s. We show by induction on k that a monic m with ms = (z - \lambda_1)\cdots(z - \lambda_k) is a product of some of the factors. If k = 0, then ms = 1 forces \deg m = 0, so m = 1, the empty product. Let k \ge 1. If m(\lambda_k) = 0, the The remainder and factor theorems gives m = (z - \lambda_k)m_1 with m_1 monic, and cancelling z - \lambda_k (a product of nonzero polynomials is nonzero, since degrees add) leaves m_1s = (z - \lambda_1)\cdots(z - \lambda_{k-1}). If m(\lambda_k) \neq 0, then m(\lambda_k)s(\lambda_k) = 0 forces s(\lambda_k) = 0, so s = (z - \lambda_k)s_1 and ms_1 = (z - \lambda_1)\cdots(z - \lambda_{k-1}). Either way the induction hypothesis finishes.
Result 3. (a) If \lambda is an eigenvalue, the Intuition shows m_T(\lambda) = 0. Conversely, let m_T(\lambda) = 0. By the The remainder and factor theorems, m_T(z) = (z - \lambda)q(z) with q monic of degree \deg m_T - 1, so q(T) \neq 0 by minimality. Pick \mathbf{v} with \mathbf{w} = q(T)\mathbf{v} \neq \mathbf{0}. Then
so \mathbf{w} is an eigenvector for \lambda. (b) As V \neq \{\mathbf{0}\}, \deg m_T \ge 1, so by the The fundamental theorem of algebra m_T has a complex zero, an eigenvalue by (a). This is a second proof of Result 1 of Upper-triangular matrices over the complex numbers.
Result 4. By Result 3(a), m_T(0) = 0 exactly when 0 is an eigenvalue, which happens exactly when T is not invertible (Result 3 of Invariant subspaces with \lambda = 0). If c_0 \neq 0, move c_0I across in m_T(T) = 0:
The two factors commute, so dividing by -c_0 gives an inverse on both sides.
Result 5. For \mathbf{u} \in U, invariance gives T^j\mathbf{u} = (T|_U)^j\mathbf{u}, so p(T|_U)\mathbf{u} = p(T)\mathbf{u} for every polynomial p. In particular m_T(T|_U) = 0, and Result 2(a) for T|_U gives m_{T|_U} \mid m_T.
Result 6. (a) Let \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n), V_k = \operatorname{span}\{\mathbf{v}_1, \dots, \mathbf{v}_k\} and q_k = (T - \lambda_1I)\cdots(T - \lambda_kI). We show by induction that q_k vanishes on V_k; the case k = n is the claim. For k = 1, \mathbf{v}_1 is an eigenvector for \lambda_1. Let k \ge 2. For j < k, \mathbf{v}_j \in V_{k-1} and the factors commute, so q_k\mathbf{v}_j = (T - \lambda_kI)q_{k-1}\mathbf{v}_j = \mathbf{0}. Column k of the matrix gives T\mathbf{v}_k = \lambda_k\mathbf{v}_k + \mathbf{u} with \mathbf{u} \in V_{k-1}, so q_k\mathbf{v}_k = q_{k-1}(T - \lambda_kI)\mathbf{v}_k = q_{k-1}\mathbf{u} = \mathbf{0}.
(b) If such a basis exists, (a) and Result 2(b) make m_T a product of linear factors. Conversely, suppose m_T(z) = (z - \mu_1)\cdots(z - \mu_r); we use induction on n = \dim V, over all operators with this property. If n = 0, the empty basis works. Let n \ge 1. Then r \ge 1, and \mu_1 is an eigenvalue by Result 3(a); extend an eigenvector \mathbf{v}_1 for it to a basis \mathcal{B} of V. Column 1 of [T]_{\mathcal{B}} is \mu_1\mathbf{e}_1, so
with Y of size (n - 1) \times (n - 1). By fact (ii), m_T(Y) is the lower-right block of m_T([T]_{\mathcal{B}}) = 0, so m_T annihilates S\colon \mathbf{x} \mapsto Y\mathbf{x} on \mathbb{F}^{n-1}, and by Result 2(b) m_S is a product of linear factors. The induction hypothesis gives a basis of \mathbb{F}^{n-1} in which S is upper triangular; if P is the invertible matrix with these vectors as columns, P^{-1}YP is upper triangular (Result 5 of The matrix of a linear map). With Q = \begin{bmatrix} 1 & \mathbf{0}^{T} \\ \mathbf{0} & P \end{bmatrix},
is upper triangular. The vectors whose \mathcal{B}-coordinates are the columns of the invertible Q form a basis \mathcal{B}' of V (Result 1 of The matrix of a linear map), and [T]_{\mathcal{B}'} = Q^{-1}[T]_{\mathcal{B}}Q by Result 5 there.
Over \mathbb{C} every monic polynomial is a product of linear factors (The fundamental theorem of algebra), so (b) recovers the upper-triangular theorem.
Worked examples
Same eigenvalues, different minimal polynomials
Find the minimal polynomials of
A monic z - c annihilates a matrix only if the matrix is cI_3, so neither minimal polynomial has degree 1. Squaring,
so m_A(z) = z^2 - z - 2 = (z - 2)(z + 1). For C,
If C^2 = aI_3 + bC, entry (1, 2) gives b = -2 and entry (1, 1) gives 3 = a + 4, so a = -1; then entry (3, 3) would be a + 2b = -5, not 4. So (I_3, C, C^2) is independent, and by Result 1 the next power must depend:
So m_C(z) = z^3 - 3z - 2 = (z - 2)(z + 1)^2. By Result 3 both matrices have exactly the eigenvalues 2 and -1, and both have characteristic polynomial -(\lambda - 2)(\lambda + 1)^2 (The characteristic equation and finding eigenvalues), yet the minimal polynomials differ. The square on z + 1 signals too few eigenvectors (When an operator is diagonalizable).
An inverse and a tenth power
With A and C as in Example 1, write C^{-1} as a polynomial in C, and write A^{10} as aA + bI_3.
Solutionm_C(0) = -2 \neq 0, so C is invertible by Result 4. From C^3 - 3C - 2I_3 = 0 we get C(C^2 - 3I_3) = 2I_3, so
For the power, divide: z^{10} = m_A(z)s(z) + az + b. Since m_A(A) = 0, A^{10} = aA + bI_3. Evaluating the division at the zeros 2 and -1 of m_A gives 2a + b = 1024 and -a + b = 1, so a = 341 and b = 342:
The shift on cubics
Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), (Tp)(x) = p(x + 1). Find m_T and every eigenvalue of T.
SolutionLet \Delta = T - I, so (\Delta p)(x) = p(x + 1) - p(x). For k \ge 1, (x + 1)^k - x^k = kx^{k-1} + (\text{lower terms}), so \Delta lowers the degree of a nonconstant polynomial by exactly one and sends constants to 0. Four applications kill every cubic, so (T - I)^4 = 0, and by Result 2(b) m_T = (z - 1)^j for some j \le 4. On x^3,
so (T - I)^3 \neq 0 and m_T(z) = (z - 1)^4. By Result 3 the only eigenvalue is 1, and E_1 = \ker\Delta is the constants. Here \deg m_T = 4 = \dim \mathcal{P}_3(\mathbb{R}), the most Result 1 allows.
The turning tray at a dim sum table
Leah and Omar share a round dim sum table with six places and a glass tray that turns. Omar records the dumplings in front of each place as \mathbf{x} = (x_1, \dots, x_6) \in \mathbb{R}^6, and one turn moves every dish to the next place:
Find m_S and the real eigenvalues, and decide whether some basis of \mathbb{R}^6 makes [S] upper triangular.
SolutionSix turns bring every dish home, so S^6 = I. On the other hand S^j\mathbf{e}_1 = \mathbf{e}_{j+1} for j = 0, \dots, 5, so if p(z) = a_0 + a_1z + \dots + a_5z^5 annihilates S, then \mathbf{0} = p(S)\mathbf{e}_1 = (a_0, a_1, \dots, a_5) and p = 0. No nonzero polynomial of degree below 6 annihilates S, so m_S(z) = z^6 - 1. Over \mathbb{R},
and both quadratics have discriminant -3, so the real zeros are 1 and -1. By Result 3 these are the real eigenvalues: a turn fixes (1, 1, 1, 1, 1, 1) and negates (1, -1, 1, -1, 1, -1).
A triangular basis would need, by Result 6(b), z^6 - 1 = (z - \mu_1)\cdots(z - \mu_6) with every \mu_j real, so every complex zero of z^6 - 1 would be real. But \displaystyle \omega = \tfrac{1}{2} + \tfrac{\sqrt{3}}{2}i is a zero of z^2 - z + 1. So no basis of \mathbb{R}^6 works. On \mathbb{C}^6, all six sixth roots of unity are eigenvalues (Complex numbers).
Matrices of one operator share its minimal polynomial
Prove that m_T = m_{[T]_{\mathcal{B}}} for every basis \mathcal{B} of V, and that similar matrices have the same minimal polynomial.
SolutionThe map R \mapsto [R]_{\mathcal{B}} on \mathcal{L}(V) is linear, sends I to I_n and composition to product (The matrix of a linear map). By induction [T^j]_{\mathcal{B}} = ([T]_{\mathcal{B}})^j, and by linearity [p(T)]_{\mathcal{B}} = p([T]_{\mathcal{B}}) for every polynomial p. An operator is 0 exactly when its matrix is the zero matrix, so p(T) = 0 if and only if p([T]_{\mathcal{B}}) = 0. The operator and its matrix have the same annihilating polynomials, hence the same monic one of least degree.
For similar matrices D = P^{-1}AP, the inner factors cancel in (P^{-1}AP)^j = P^{-1}A^jP, so p(D) = P^{-1}p(A)P for every p. Since P is invertible, this is 0 exactly when p(A) = 0, and m_D = m_A.
Reading a triangular matrix
Find the minimal polynomial of
M is upper triangular with diagonal 1, 3, 1, 1. By Result 6(a), (M - I_4)^3(M - 3I_4) = 0, so by Result 2(b) m_M is a product of some of these four factors. The eigenvalues 1 and 3 (Result 4 of Upper-triangular matrices over the complex numbers) are zeros of m_M by Result 3, which leaves (z - 1)(z - 3), (z - 1)^2(z - 3) and (z - 1)^3(z - 3). The first fails:
This product has only a fourth column, (6, 0, -4, 0), and each row of M - I_4 has dot product 0 with it, so (M - I_4)^2(M - 3I_4) = 0. By Result 2(a), m_M(z) = (z - 1)^2(z - 3).
Boundaries
- Over \mathbb{R} a minimal polynomial may have no zeros. The quarter turn (x, y) \mapsto (-y, x) squares to -I and is not a multiple of I, so its minimal polynomial is z^2 + 1. By Result 3(a) it has no real eigenvalue.
- Finite dimension is needed. On \mathcal{P}(\mathbb{F}), differentiation D is annihilated by no nonzero polynomial: if a_j is the lowest nonzero coefficient of p, then p(D)x^j = j!\,a_j \neq 0, because D^ix^j = 0 for i > j.
- The eigenvalues do not determine m_T (Example 1), and m_T does not determine T up to similarity. \operatorname{diag}(1, 1, 2) and \operatorname{diag}(1, 2, 2) both have minimal polynomial (z - 1)(z - 2), but \dim E_1 is 2 for one and 1 for the other, so they are not similar. Jordan canonical form gives the complete answer.
Where this goes next
When an operator is diagonalizable proves that T is diagonalizable exactly when m_T is a product of distinct linear factors, as for A in Example 1 and not for C. Generalized eigenvectors proves the Cayley-Hamilton theorem, so m_T divides the characteristic polynomial, and Jordan canonical form reads the exponent of each z - \lambda in m_T as the size of the largest Jordan block for \lambda.
Exercises
Computing minimal polynomials
- Find m_A for A = \begin{bmatrix} 4 & -1 \\ 2 & 1 \end{bmatrix}, the eigenvalues of A, and A^{-1} as a polynomial in A.
- Find m_B for the 3 \times 3 matrix B whose every row is (1, 2, 3). Is B invertible?
- Find m_A for A = \begin{bmatrix} 3 & -1 & 2 \\ 1 & 1 & 2 \\ 0 & 0 & 2 \end{bmatrix}.
- Let T \in \mathcal{L}(\mathcal{P}_2(\mathbb{R})), (Tp)(x) = p(2x). Find m_T.
- Let T \in \mathcal{L}(\mathbb{R}^{2 \times 2}), TX = X^{T}. Find m_T and a basis of each eigenspace.
Using the minimal polynomial
- An operator T on a real vector space has m_T(z) = z^3 - 2z + 1. (a) Find the eigenvalues of T. (b) Write T^{-1} and T^4 as polynomials in T of degree at most 2.
- For A of Exercise 1, write A^6 as aA + bI_2.
- Let T \in \mathcal{L}(\mathbb{R}^4) have m_T(z) = (z^2 + 1)(z - 2). Find the eigenvalues of T, write T^{-1} as a polynomial in T, and decide whether some basis of \mathbb{R}^4 gives T an upper-triangular matrix.
- Omar’s next table has four places, and its tray turns two places at a time: R(x_1, x_2, x_3, x_4) = (x_3, x_4, x_1, x_2). Find m_R, the eigenvalues of R and a basis of each eigenspace.
Proofs
- Let V \neq \{\mathbf{0}\}. Prove that \deg m_T = 1 if and only if T = \lambda I for some \lambda \in \mathbb{F}.
- (a) Prove that if p(T) = 0, then p(\lambda) = 0 for every eigenvalue \lambda of T. (b) Refute with a counterexample: if p(\lambda) = 0 for every eigenvalue \lambda of T, then p(T) = 0.
- Let d = \deg m_T. Prove that (I, T, \dots, T^{d-1}) is a basis of \operatorname{span}\{I, T, T^2, \dots\}, so this subspace of \mathcal{L}(V) has dimension \deg m_T.
- Suppose V = U \oplus W with U and W invariant under T. Prove that q(T) = 0 if and only if m_{T|_U} \mid q and m_{T|_W} \mid q, and deduce that m_T \mid m_{T|_U}m_{T|_W}.
- Let T be invertible and d = \deg m_T. Prove that m_{T^{-1}}(z) = z^dm_T(1/z)/m_T(0).
- Prove that if T^k = 0 for some k \ge 1, then m_T(z) = z^j with j \le \dim V, and hence T^{\dim V} = 0.
Answers
- A is not a multiple of I_2, and A^2 = \begin{bmatrix} 14 & -5 \\ 10 & -1 \end{bmatrix} = 5A - 6I_2, so m_A(z) = z^2 - 5z + 6 = (z - 2)(z - 3), with eigenvalues 2 and 3 (Result 3). From A(5I_2 - A) = 6I_2, A^{-1} = \tfrac{1}{6}(5I_2 - A) = \tfrac{1}{6}\begin{bmatrix} 1 & 1 \\ -2 & 4 \end{bmatrix}.
- Column j of B has every entry j, so entry (i, j) of B^2 is (1 + 2 + 3)j = 6j, and B^2 = 6B. As B \neq cI_3, m_B(z) = z^2 - 6z = z(z - 6), with constant term 0: B is not invertible (Result 4).
- A - 2I_3 \neq 0 has rows (1, -1, 2), (1, -1, 2), \mathbf{0}, and its square is 0, since (1, -1, 2) has dot product 0 with every column. So (z - 2)^2 annihilates A and z - 2 does not, and by Result 2(b) m_A(z) = (z - 2)^2.
- T1 = 1, Tx = 2x, Tx^2 = 4x^2, so the matrix is \operatorname{diag}(1, 2, 4) and Result 6(a) gives (T - I)(T - 2I)(T - 4I) = 0. By Results 2(b) and 3, m_T contains each factor: m_T(z) = (z - 1)(z - 2)(z - 4).
- Let E_{ij} have 1 in entry (i, j) and 0 elsewhere. T^2X = X, so z^2 - 1 annihilates T, and T \neq \pm I since T swaps E_{12} and E_{21}. So m_T(z) = z^2 - 1. E_1 is the symmetric matrices, basis E_{11}, E_{22}, E_{12} + E_{21}; E_{-1} is the matrices with X^{T} = -X, basis E_{12} - E_{21}.
- (a) z^3 - 2z + 1 = (z - 1)(z^2 + z - 1), with zeros 1 and \displaystyle \tfrac{-1 \pm \sqrt{5}}{2}, all real, so these three are the eigenvalues (Result 3). (b) T^3 - 2T = -I gives T(2I - T^2) = I, so T^{-1} = 2I - T^2, and T^4 = T(2T - I) = 2T^2 - T.
- z^6 = m_A(z)s(z) + az + b; at 2 and 3, 2a + b = 64 and 3a + b = 729, so a = 665, b = -1266 and A^6 = 665A - 1266I_2.
- The only real zero of m_T is 2, the only eigenvalue. m_T(z) = z^3 - 2z^2 + z - 2 gives T(T^2 - 2T + I) = 2I, so T^{-1} = \tfrac{1}{2}(T^2 - 2T + I). No basis works: a product of real linear factors has only real zeros, but m_T(i) = 0, so Result 6(b) fails.
- R^2 = I and R \neq \pm I (as R\mathbf{e}_1 = \mathbf{e}_3), so m_R(z) = z^2 - 1, with eigenvalues 1 and -1. E_1 = \{\mathbf{x} : x_1 = x_3,\ x_2 = x_4\}, basis (1, 0, 1, 0), (0, 1, 0, 1); E_{-1} has basis (1, 0, -1, 0), (0, 1, 0, -1).
- If T = \lambda I, then z - \lambda annihilates T, and degree 0 is impossible on V \neq \{\mathbf{0}\}, so m_T(z) = z - \lambda. Conversely, if m_T(z) = z - \lambda, then T - \lambda I = m_T(T) = 0.
- (a) If T\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} \neq \mathbf{0}, then \mathbf{0} = p(T)\mathbf{v} = p(\lambda)\mathbf{v} (as in the Intuition), so p(\lambda) = 0. (b) Let T(x, y) = (y, 0) on \mathbb{R}^2. T^2 = 0, so by (a) every eigenvalue is a zero of z^2, and T\mathbf{e}_1 = \mathbf{0} makes 0 an eigenvalue. So p(z) = z vanishes at every eigenvalue, yet p(T) = T \neq 0.
- Independence: a relation a_0I + \dots + a_{d-1}T^{d-1} = 0 with some a_j \neq 0 gives a nonzero annihilator of degree below d, monic after dividing by its leading coefficient, against minimality. Spanning: for k \ge 0, divide z^k = m_Ts + r with r = 0 or \deg r < d; then T^k = r(T), a combination of I, \dots, T^{d-1}.
- Write \mathbf{v} = \mathbf{u} + \mathbf{w} with \mathbf{u} \in U, \mathbf{w} \in W. As in the proof of Result 5, q(T)\mathbf{v} = q(T|_U)\mathbf{u} + q(T|_W)\mathbf{w}. If both minimal polynomials divide q, both terms are \mathbf{0} by Result 2(a) for the restrictions, so q(T) = 0. If q(T) = 0, then q(T|_U)\mathbf{u} = q(T)\mathbf{u} = \mathbf{0} for every \mathbf{u} \in U, so m_{T|_U} \mid q, and likewise for W. Both divide q = m_{T|_U}m_{T|_W}, so q(T) = 0 and m_T \mid q by Result 2(a).
- Write m_T(z) = \sum_{j=0}^d c_jz^j with c_d = 1 and c_0 \neq 0 (Result 4), and q(z) = z^dm_T(1/z)/c_0 = \sum_{j=0}^d (c_j/c_0)z^{d-j}, monic of degree d. Then q(T^{-1}) = \tfrac{1}{c_0}T^{-d}\sum_j c_jT^j = \tfrac{1}{c_0}T^{-d}m_T(T) = 0, so m_{T^{-1}} \mid q and \deg m_{T^{-1}} \le d. The same argument for T^{-1} gives d \le \deg m_{T^{-1}}, and a monic divisor of q of degree d is q.
- z^k = (z - 0)\cdots(z - 0) annihilates T, so by Result 2(b) m_T(z) = z^j, and j = \deg m_T \le \dim V by Result 1. Then T^{\dim V} = T^{\dim V - j}m_T(T) = 0.