Upper-triangular matrices over the complex numbers
Over the real numbers a rotation has no eigenvector, and even over the complex numbers some operators have too few eigenvectors to make a basis. What simple matrix can every complex operator be given?
Where you are 5 in Eigenvalues, proved
Why this exists
Diagonalization gives an operator its simplest matrix, a diagonal one, whenever there is a basis of eigenvectors. Often there is none. The quarter turn (x, y) \mapsto (-y, x) of \mathbb{R}^2 has no eigenvector at all, and T(x, y) = (3x + y,\ -x + y) has eigenvectors on one line only, over \mathbb{R} and over \mathbb{C} alike.
Over \mathbb{C} there is always a next best thing. In a well-chosen basis, the matrix of any operator on a nonzero finite-dimensional complex space is upper triangular, with zeros below the diagonal and the eigenvalues on it. That is close enough to diagonal to read off eigenvalues, test invertibility and compute powers.
Definition
A square matrix A = [a_{jk}] is upper triangular if every entry below the diagonal is zero: a_{jk} = 0 whenever j > k. Its diagonal is a_{11}, \dots, a_{nn}.
Let V be finite-dimensional over \mathbb{F}, let T \in \mathcal{L}(V) and let \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n) be a basis of V. We say T has an upper-triangular matrix with respect to \mathcal{B} when [T]_{\mathcal{B}} is upper triangular. For k = 0, 1, \dots, n write
so that V_0 = \{\mathbf{0}\}, V_n = V, V_0 \subseteq V_1 \subseteq \dots \subseteq V_n and \dim V_k = k.
Intuition
Column k of [T]_{\mathcal{B}} holds the coordinates of T\mathbf{v}_k (The matrix of a linear map). The matrix is upper triangular when each T\mathbf{v}_k is a multiple of \mathbf{v}_k plus a combination of earlier basis vectors. So \mathbf{v}_1 is an eigenvector, \mathbf{v}_2 is one up to an error in V_1, and so on: T never pushes a vector of V_k out of V_k.
In the figure T(x, y) = (3x + y,\ -x + y), \mathbf{v}_1 = (1, -1) and \mathbf{v}_2 = (0, 1). Then T\mathbf{v}_1 = 2\mathbf{v}_1 and T\mathbf{v}_2 = (1, 1) = \mathbf{v}_1 + 2\mathbf{v}_2, so
Results
Throughout, V is finite-dimensional and T \in \mathcal{L}(V). Results 2, 4 and 5 hold over any field with the same proofs; Results 1 and 3 need \mathbb{F} = \mathbb{C}.
Result 1 (eigenvalues exist over C). If V is a nonzero finite-dimensional complex vector space, every T \in \mathcal{L}(V) has an eigenvalue.
Result 2 (conditions for an upper-triangular matrix). Let \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n) be a basis of V. The following are equivalent: (a) [T]_{\mathcal{B}} is upper triangular; (b) T\mathbf{v}_k \in V_k for k = 1, \dots, n; (c) V_k is invariant under T for k = 1, \dots, n.
Result 3 (upper-triangular form). If V is a nonzero finite-dimensional complex vector space and T \in \mathcal{L}(V), then V has a basis with respect to which T has an upper-triangular matrix.
Result 4 (the diagonal). Suppose [T]_{\mathcal{B}} is upper triangular with diagonal \lambda_1, \dots, \lambda_n. Then (a) T is invertible if and only if every \lambda_k \neq 0; (b) the eigenvalues of T are exactly \lambda_1, \dots, \lambda_n.
Result 5 (polynomials of a triangular operator). Suppose [T]_{\mathcal{B}} is upper triangular with diagonal \lambda_1, \dots, \lambda_n. For every polynomial p with coefficients in \mathbb{F}, [p(T)]_{\mathcal{B}} is upper triangular with diagonal p(\lambda_1), \dots, p(\lambda_n), so the eigenvalues of p(T) are exactly these numbers. In particular, if V \neq \{\mathbf{0}\} is complex, the eigenvalues of p(T) are exactly the numbers p(\lambda) with \lambda an eigenvalue of T.
Why it is true
Result 1. Let n = \dim V \ge 1 and pick \mathbf{v} \neq \mathbf{0}. The list \mathbf{v}, T\mathbf{v}, \dots, T^n\mathbf{v} has n + 1 vectors in an n-dimensional space, so it is linearly dependent (Bases and dimension, proved): a_0\mathbf{v} + a_1T\mathbf{v} + \dots + a_nT^n\mathbf{v} = \mathbf{0} with not every a_j zero. Not all of a_1, \dots, a_n are zero, since otherwise a_0\mathbf{v} = \mathbf{0} forces a_0 = 0 too. So p(z) = a_0 + a_1z + \dots + a_nz^n has some degree m \ge 1, and by the The fundamental theorem of algebra and the The remainder and factor theorems it splits as p(z) = c(z - \lambda_1)\cdots(z - \lambda_m) with c \neq 0. Products of polynomials become compositions of operators (Result 5 of Invariant subspaces), so
A composition of injective maps is injective, and this one sends \mathbf{v} \neq \mathbf{0} to \mathbf{0}. So some factor T - \lambda_j I is not injective, and \lambda_j is an eigenvalue (Result 3 of Invariant subspaces).
Result 2. (a) to (b): column k is zero below row k, so T\mathbf{v}_k = a_{1k}\mathbf{v}_1 + \dots + a_{kk}\mathbf{v}_k \in V_k. (b) to (c): for j \le k, T\mathbf{v}_j \in V_j \subseteq V_k; these \mathbf{v}_j span V_k and T is linear, so T maps V_k into V_k. (c) to (a): T\mathbf{v}_k \in V_k is a combination of \mathbf{v}_1, \dots, \mathbf{v}_k, and coordinates are unique (Result 1 of Bases and dimension, proved), so column k is zero below row k.
Result 3. Induction on n = \dim V (Strong induction and well-ordering). If n = 1, every 1 \times 1 matrix is upper triangular. Let n \ge 2 and assume the result for every nonzero complex space of smaller dimension. By Result 1, T has an eigenvalue \lambda. Let
U is invariant under T (Result 6 of Invariant subspaces), and T - \lambda I is not surjective (Result 3 there), so m = \dim U < n. If m \ge 1, the induction hypothesis applied to T|_U \in \mathcal{L}(U) gives a basis (\mathbf{u}_1, \dots, \mathbf{u}_m) of U with T\mathbf{u}_j \in \operatorname{span}\{\mathbf{u}_1, \dots, \mathbf{u}_j\} for each j, by Result 2; if m = 0, take the empty list. Extend it to a basis (\mathbf{u}_1, \dots, \mathbf{u}_m, \mathbf{w}_1, \dots, \mathbf{w}_r) of V (Bases and dimension, proved). For each k,
and the first term lies in U. So T\mathbf{w}_k \in \operatorname{span}\{\mathbf{u}_1, \dots, \mathbf{u}_m, \mathbf{w}_k\}, inside the span of all basis vectors up to \mathbf{w}_k. Every basis vector satisfies (b) of Result 2, so the matrix is upper triangular, and each \mathbf{w}_k has diagonal entry \lambda.
Result 4. (a) Suppose every \lambda_k \neq 0. We show \mathbf{v}_k \in \operatorname{im} T by induction on k. Since T\mathbf{v}_1 = \lambda_1\mathbf{v}_1, \mathbf{v}_1 = T(\lambda_1^{-1}\mathbf{v}_1). If \mathbf{v}_1, \dots, \mathbf{v}_{k-1} \in \operatorname{im} T, write T\mathbf{v}_k = \lambda_k\mathbf{v}_k + \mathbf{u} with \mathbf{u} \in V_{k-1} \subseteq \operatorname{im} T; then \mathbf{v}_k = \lambda_k^{-1}(T\mathbf{v}_k - \mathbf{u}) \in \operatorname{im} T. So T is surjective, hence invertible (Result 3 of Invariant subspaces with \lambda = 0). Conversely, suppose \lambda_k = 0. Then T\mathbf{v}_k \in V_{k-1}, and T\mathbf{v}_j \in V_j \subseteq V_{k-1} for j < k, so T maps V_k into V_{k-1}. A linear map into a space of smaller dimension is not injective, by the fundamental theorem of linear maps (Null space, range and the fundamental theorem), so T is not invertible.
(b) [T - \lambda I]_{\mathcal{B}} = [T]_{\mathcal{B}} - \lambda I_n is upper triangular with diagonal \lambda_1 - \lambda, \dots, \lambda_n - \lambda. By Result 3 of Invariant subspaces, \lambda is an eigenvalue exactly when T - \lambda I is not invertible, which by (a) happens exactly when \lambda = \lambda_k for some k.
Result 5. Let A and C be upper-triangular n \times n matrices. Entry (j, k) of AC is \sum_i a_{ji}c_{ik}, and a term can be nonzero only if j \le i \le k. For j > k no term survives, and for j = k only i = j does. So AC is upper triangular with diagonal a_{11}c_{11}, \dots, a_{nn}c_{nn}; sums and scalar multiples behave the same way, entry by entry. The matrix of a composition is the product of the matrices (The matrix of a linear map), so by induction [T^j]_{\mathcal{B}} = ([T]_{\mathcal{B}})^j is upper triangular with diagonal \lambda_1^j, \dots, \lambda_n^j, and [p(T)]_{\mathcal{B}} = \sum_j a_j([T]_{\mathcal{B}})^j has diagonal p(\lambda_1), \dots, p(\lambda_n). Result 4(b) applied to p(T) names its eigenvalues. Over \mathbb{C}, Result 3 supplies such a basis, and by Result 4(b) its diagonal lists every eigenvalue of T.
Worked examples
A first-order operator on quadratics
Let T \in \mathcal{L}(\mathcal{P}_2(\mathbb{R})), (Tp)(x) = p(x) + (x - 2)p'(x). Show that T has an upper-triangular matrix with respect to (1, x, x^2) but not (x^2, x, 1), find its eigenvalues, and find a basis of eigenvectors.
SolutionT1 = 1, Tx = x + (x - 2) = 2x - 2 and Tx^2 = x^2 + 2x(x - 2) = 3x^2 - 4x. Each T(x^k) lies in the span of 1, \dots, x^k, so by Result 2
In the order (x^2, x, 1), Tx = 2x - 2 is not in \operatorname{span}\{x^2, x\}, so (b) of Result 2 fails. By Result 4 the eigenvalues are 1, 2 and 3, found over \mathbb{R}. Centering at x = 2 gives the eigenvectors:
So in the basis (1, x - 2, (x - 2)^2) the matrix is \operatorname{diag}(1, 2, 3).
An eigenvalue from the list in the proof
Let T \in \mathcal{L}(\mathbb{C}^3), T(z_1, z_2, z_3) = (z_1 - 2z_2,\ z_1 - z_2,\ 3z_3). Starting from \mathbf{v} = \mathbf{e}_1, follow the proof of Result 1 to an eigenvalue and an eigenvector.
SolutionT\mathbf{e}_1 = (1, 1, 0) and T^2\mathbf{e}_1 = T(1, 1, 0) = (-1, 0, 0) = -\mathbf{e}_1. The list \mathbf{v}, T\mathbf{v}, T^2\mathbf{v}, T^3\mathbf{v} is dependent through T^2\mathbf{v} + \mathbf{v} = \mathbf{0}, so p(z) = z^2 + 1 = (z - i)(z + i) and
Apply the right factor first: \mathbf{w} = (T + iI)\mathbf{e}_1 = (1 + i, 1, 0) \neq \mathbf{0}, and the left factor sends \mathbf{w} to \mathbf{0}, so T\mathbf{w} = i\mathbf{w}. Directly, T\mathbf{w} = (-1 + i, i, 0), which is i(1 + i, 1, 0). The other order gives (T - iI)\mathbf{e}_1 = (1 - i, 1, 0), an eigenvector for -i.
Following the proof of Result 3
Let A = \begin{bmatrix} 1 & 1 & 0 \\ 1 & 2 & -1 \\ 0 & 1 & 1 \end{bmatrix} act on \mathbb{C}^3. Build a basis in which its matrix is upper triangular, as the proof of Result 3 does, and show that no basis makes it diagonal.
SolutionExpanding along the first row (The characteristic equation and finding eigenvalues),
Take \lambda = 2. The columns of A - 2I are (-1, 1, 0), (1, 0, 1) and (0, -1, -1), the last minus the sum of the first two, so
a plane invariant under A. Inside U the remaining eigenvalue is 1: A - I has rows (0, 1, 0), (1, 1, -1), (0, 1, 0) and kills exactly the multiples of \mathbf{u}_1 = (1, 0, 1), which lies in U. Complete it to a basis of U with \mathbf{u}_2 = (0, 1, 1):
So (\mathbf{u}_1, \mathbf{u}_2) is a triangular basis for A|_U. Extend by \mathbf{e}_1 \notin U. Then (A - 2I)\mathbf{e}_1 = (-1, 1, 0) = -\mathbf{u}_1 + \mathbf{u}_2 \in U, so A\mathbf{e}_1 = -\mathbf{u}_1 + \mathbf{u}_2 + 2\mathbf{e}_1. In the basis \mathcal{B} = (\mathbf{u}_1, \mathbf{u}_2, \mathbf{e}_1)
The eigenvalue chosen first, 2, lands last. No basis diagonalizes A: E_1 = \operatorname{span}\{(1, 0, 1)\} and E_2 = \ker(A - 2I) = \operatorname{span}\{(1, 1, 1)\} are two lines, too few eigenvectors for a basis (Diagonalization).
Zero diagonal forces a zero power
Let V be an n-dimensional space over any field and suppose [T]_{\mathcal{B}} is upper triangular with every diagonal entry 0. Prove that T^n = 0. Deduce that if V \neq \{\mathbf{0}\} is complex and 0 is the only eigenvalue of T, then T^{\dim V} = 0.
SolutionA zero diagonal entry in column k means T\mathbf{v}_k \in V_{k-1}. For j \le k, T\mathbf{v}_j \in V_{j-1} \subseteq V_{k-1}, and these \mathbf{v}_j span V_k, so by linearity T(V_k) \subseteq V_{k-1} for k = 1, \dots, n. Each application of T lowers the index by one:
So T^n = 0. For the second claim, Result 3 gives a basis with [T]_{\mathcal{B}} upper triangular, and by Result 4(b) its diagonal entries are eigenvalues of T, so all are 0.
The eigenvalues of a polynomial in T
Let V be a complex vector space of dimension 4 and let T \in \mathcal{L}(V) have exactly the eigenvalues 1 and i. Find the eigenvalues of T^2 + T, decide whether T^2 + I is invertible, and prove that (T^4 - I)^4 = 0.
SolutionBy Result 3 choose \mathcal{B} with [T]_{\mathcal{B}} upper triangular. By Result 4(b) each diagonal entry is 1 or i, and both occur.
For p(z) = z^2 + z, p(1) = 2 and p(i) = -1 + i, so by Result 5 the eigenvalues of T^2 + T are 2 and -1 + i. For q(z) = z^2 + 1, q(i) = 0 sits on the diagonal of [q(T)]_{\mathcal{B}}, so T^2 + I is not invertible by Result 4(a). Finally r(z) = z^4 - 1 has r(1) = r(i) = 0, so [T^4 - I]_{\mathcal{B}} is upper triangular with zero diagonal, and Example 4 gives (T^4 - I)^4 = 0.
The fourth power cannot be dropped: if [T]_{\mathcal{B}} has diagonal blocks \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix} and \begin{bmatrix} i & 1 \\ 0 & i \end{bmatrix}, the top block of [T^4]_{\mathcal{B}} is \begin{bmatrix} 1 & 4 \\ 0 & 1 \end{bmatrix}, so T^4 \neq I.
The geocache puzzle
Amara hides a geocache and posts a puzzle for Hiro: start from the pair (1, 0) and replace (x, y) by (3x + y,\ -x + y) ten times; the last pair opens the lock. The first pairs are (3, -1), (8, -4), (20, -12). Hiro wants a formula for every step.
SolutionThe step is the operator of the figure, A = \begin{bmatrix} 3 & 1 \\ -1 & 1 \end{bmatrix}, with \det(A - \lambda I) = \lambda^2 - 4\lambda + 4 = (\lambda - 2)^2. Its only eigenvalue is 2 and E_2 = \operatorname{span}\{(1, -1)\} is a line, so Diagonalization is not available. In the triangular basis \mathcal{B} = ((1, -1), (0, 1)),
and N^2 = 0.
By induction (2I_2 + N)^k = 2^kI_2 + k2^{k-1}N: multiplying by 2I_2 + N gives 2^{k+1}I_2 + 2^kN + k2^kN, since N^2 = 0. The start (1, 0) = (1, -1) + (0, 1) has coordinates (1, 1), so after k steps the coordinates are (2^k + k2^{k-1},\ 2^k) and the pair is
For k = 1, 2, 3 this matches the puzzle, and for k = 10 the lock opens at (6{,}144,\ -5{,}120).
Boundaries
- Over \mathbb{R} the theorem fails. The first vector of a triangular basis is an eigenvector, so a quarter turn of \mathbb{R}^2 has no upper-triangular matrix over \mathbb{R}. The minimal polynomial gives the exact test over any field.
- Finite dimension is needed. On the space of complex sequences, the right shift S(a_1, a_2, \dots) = (0, a_1, a_2, \dots) has no eigenvalue: S\mathbf{a} = \lambda\mathbf{a} gives \lambda a_1 = 0 and a_k = \lambda a_{k+1}, which force every a_k = 0 whether \lambda = 0 or not.
- Triangular is not diagonal. The operators of Examples 3 and 6 have no basis of eigenvectors. When an operator is diagonalizable says when the entries above the diagonal can be cleared, and Jordan canonical form gives the sparsest pattern that always exists.
- The basis and the order of the diagonal are not unique (Exercise 12), though the number of times each eigenvalue appears is the same for every triangular basis, as Generalized eigenvectors shows. Only \mathbf{v}_1 need be an eigenvector (Exercise 16).
Where this goes next
The minimal polynomial studies the polynomials that send T to 0, among them (T - \lambda_1 I)\cdots(T - \lambda_n I) for the diagonal entries of a triangular matrix. The The spectral theorem and Jordan canonical form refine the picture, and Trace and determinant from the structure reads the trace and determinant off the diagonal.
Exercises
Reading an upper-triangular matrix
- Let T(x_1, x_2, x_3) = (2x_1 + x_3,\ x_1 + 3x_2,\ 2x_3) on \mathbb{R}^3. With respect to which bases does T have an upper-triangular matrix: (a) (\mathbf{e}_1, \mathbf{e}_2, \mathbf{e}_3), (b) (\mathbf{e}_2, \mathbf{e}_1, \mathbf{e}_3), (c) ((1, -1, 0), \mathbf{e}_2, \mathbf{e}_3), (d) (\mathbf{e}_3, \mathbf{e}_2, \mathbf{e}_1)? Give the matrix in each case that works.
- Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), (Tp)(x) = xp'(x) + p(x + 1). Find [T] in the basis (1, x, x^2, x^3), the eigenvalues of T, and an eigenvector for 2.
- Let T \in \mathcal{L}(\mathbb{C}^3) have matrix \begin{bmatrix} 2 & 5 & 1 \\ 0 & -1 & 3 \\ 0 & 0 & i \end{bmatrix} in some basis. (a) Find the eigenvalues of T^2 - T and of T^3. (b) Which of T^2 + T + I and T^2 - T - 2I is invertible?
- Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{C})), Tp = p' + p''. Show that 0 is the only eigenvalue, that T^4 = 0 and that T^3 \neq 0, and find the eigenspace E_0.
An eigenvalue from a list
- Let A = \begin{bmatrix} 3 & -2 \\ 1 & 1 \end{bmatrix} act on \mathbb{C}^2. From \mathbf{v} = \mathbf{e}_1 find a linear dependence among \mathbf{v}, A\mathbf{v}, A^2\mathbf{v}, factor the polynomial, and produce an eigenvector for each eigenvalue.
- Let D \in \mathcal{L}(\mathcal{P}_2(\mathbb{C})), Dp = p'. Run the proof of Result 1 from \mathbf{v} = x^2.
Finding a triangular basis
- Let A = \begin{bmatrix} 5 & -4 \\ 1 & 1 \end{bmatrix}. Find a basis of \mathbb{C}^2 in which its matrix is upper triangular, and show that no basis makes it diagonal.
- Let A = \begin{bmatrix} 3 & 1 & 0 \\ 0 & 3 & 0 \\ 4 & -3 & -1 \end{bmatrix}. Follow the proof of Result 3 with \lambda = -1: find U = \operatorname{im}(A + I), a triangular basis of A|_U, and the extended basis with the matrix of A in it.
- Let T(x, y) = (x - 2y,\ 2x + y). Show that no basis of \mathbb{R}^2 gives T an upper-triangular matrix, and find a basis of \mathbb{C}^2 that does.
- Hiro answers Amara with a cache of his own: start from (1, 0) and replace (x, y) by (5x - 4y,\ x + y). Using Exercise 7, find the pair after k steps and the pair after 6 steps.
Proofs
- Prove that [T]_{(\mathbf{v}_1, \dots, \mathbf{v}_n)} is upper triangular if and only if [T]_{(\mathbf{v}_n, \dots, \mathbf{v}_1)} is lower triangular (zero above the diagonal).
- Let V \neq \{\mathbf{0}\} be finite-dimensional and complex and \lambda an eigenvalue of T. Prove that some basis gives T an upper-triangular matrix whose last diagonal entry is \lambda.
- Suppose [T]_{\mathcal{B}} is upper triangular and the eigenvalue \lambda appears exactly k times on its diagonal. Prove that \dim E_\lambda \le k.
- Let V \neq \{\mathbf{0}\} be finite-dimensional and complex and S, T \in \mathcal{L}(V) with ST = TS. Prove that S and T have a common eigenvector.
- Prove, over any field, that if T^m = 0 for some m \ge 1, then every eigenvalue of T is 0. With Example 4, conclude that on a nonzero finite-dimensional complex space, T^m = 0 for some m exactly when 0 is the only eigenvalue of T.
- Refute with a counterexample: if [T]_{(\mathbf{v}_1, \dots, \mathbf{v}_n)} is upper triangular, then every \mathbf{v}_k is an eigenvector of T.
Answers
- (a) No: T\mathbf{e}_1 = (2, 1, 0) \notin \operatorname{span}\{\mathbf{e}_1\}. (b) Yes: T\mathbf{e}_2 = 3\mathbf{e}_2, T\mathbf{e}_1 = \mathbf{e}_2 + 2\mathbf{e}_1, T\mathbf{e}_3 = \mathbf{e}_1 + 2\mathbf{e}_3, matrix \begin{bmatrix} 3 & 1 & 0 \\ 0 & 2 & 1 \\ 0 & 0 & 2 \end{bmatrix}. (c) Yes: T(1, -1, 0) = 2(1, -1, 0), T\mathbf{e}_2 = 3\mathbf{e}_2, T\mathbf{e}_3 = (1, 0, 2) = (1, -1, 0) + \mathbf{e}_2 + 2\mathbf{e}_3, matrix \begin{bmatrix} 2 & 0 & 1 \\ 0 & 3 & 1 \\ 0 & 0 & 2 \end{bmatrix}. (d) No: T\mathbf{e}_3 \notin \operatorname{span}\{\mathbf{e}_3\}. By Result 4 the eigenvalues are 2 and 3.
- T1 = 1, Tx = 2x + 1, Tx^2 = 3x^2 + 2x + 1, Tx^3 = 4x^3 + 3x^2 + 3x + 1, so [T] = \begin{bmatrix} 1 & 1 & 1 & 1 \\ 0 & 2 & 2 & 3 \\ 0 & 0 & 3 & 3 \\ 0 & 0 & 0 & 4 \end{bmatrix}, with eigenvalues 1, 2, 3, 4 (Result 4). T(x + 1) = (2x + 1) + 1 = 2(x + 1).
- (a) By Result 5, T^2 - T has diagonal 2, 2, -1 - i: eigenvalues 2 and -1 - i. T^3 has eigenvalues 8, -1, -i. (b) z^2 + z + 1 takes the values 7, 1, i, all nonzero, so T^2 + T + I is invertible. z^2 - z - 2 = (z - 2)(z + 1) vanishes at 2, so T^2 - T - 2I is not.
- T1 = 0, Tx = 1, Tx^2 = 2x + 2, Tx^3 = 3x^2 + 6x: the matrix in (1, x, x^2, x^3) is upper triangular with zero diagonal, so 0 is the only eigenvalue (Result 4) and T^4 = 0 (Example 4). T^3x^3 = T^2(3x^2 + 6x) = T(6x + 12) = 6 \neq 0. T(a + bx + cx^2 + dx^3) = (b + 2c) + (2c + 6d)x + 3dx^2 is 0 only if b = c = d = 0, so E_0 = \operatorname{span}\{1\}.
- A\mathbf{v} = (3, 1) and A^2\mathbf{v} = (7, 4) = 4A\mathbf{v} - 5\mathbf{v}, so p(z) = z^2 - 4z + 5 = (z - 2 - i)(z - 2 + i). (A - (2 - i)I)\mathbf{e}_1 = (1 + i, 1) \ne \mathbf{0}, and A(1 + i, 1) = (1 + 3i, 2 + i) = (2 + i)(1 + i, 1). Since A is real, conjugating gives eigenvalue 2 - i with eigenvector (1 - i, 1).
- D\mathbf{v} = 2x, D^2\mathbf{v} = 2, D^3\mathbf{v} = 0, so p(z) = z^3, three factors D - 0I. Applying them in turn gives 2x, then 2, then 0, so 0 is an eigenvalue with eigenvector the constant 2.
- \det(A - \lambda I) = \lambda^2 - 6\lambda + 9 = (\lambda - 3)^2, and A - 3I = \begin{bmatrix} 2 & -4 \\ 1 & -2 \end{bmatrix} has kernel \operatorname{span}\{(2, 1)\}. In the basis ((2, 1), \mathbf{e}_1), A(2, 1) = 3(2, 1) and A\mathbf{e}_1 = (5, 1) = (2, 1) + 3\mathbf{e}_1, giving \begin{bmatrix} 3 & 1 \\ 0 & 3 \end{bmatrix}. A diagonal basis would consist of eigenvectors, all on the line E_3, so none exists.
- A + I has columns (4, 0, 4), (1, 4, -3), \mathbf{0}, so U = \operatorname{span}\{(1, 0, 1), (1, 4, -3)\}, the plane \{\mathbf{x} : x_1 = x_2 + x_3\}. In U: A(1, 0, 1) = 3(1, 0, 1) and A(0, 1, -1) = (1, 3, -2) = (1, 0, 1) + 3(0, 1, -1). Extend by \mathbf{e}_3 \notin U, with A\mathbf{e}_3 = -\mathbf{e}_3. In ((1, 0, 1), (0, 1, -1), \mathbf{e}_3) the matrix is \begin{bmatrix} 3 & 1 & 0 \\ 0 & 3 & 0 \\ 0 & 0 & -1 \end{bmatrix}.
- \det(A - \lambda I) = \lambda^2 - 2\lambda + 5 = (\lambda - 1)^2 + 4 > 0 for real \lambda, so T has no real eigenvalue, while the first vector of a triangular basis is an eigenvector. Over \mathbb{C}, T(1, -i) = (1 + 2i)(1, -i) and T(1, i) = (1 - 2i)(1, i), so ((1, -i), (1, i)) gives \operatorname{diag}(1 + 2i, 1 - 2i).
- In \mathcal{B} = ((2, 1), \mathbf{e}_1), [A]_{\mathcal{B}} = 3I_2 + N with N^2 = 0, so [A^k]_{\mathcal{B}} = 3^kI_2 + k3^{k-1}N as in Example 6. The start (1, 0) has coordinates (0, 1), which go to (k3^{k-1}, 3^k), so the pair is k3^{k-1}(2, 1) + 3^k(1, 0) = \big(3^{k-1}(2k + 3),\ k3^{k-1}\big). For k = 6: (3645, 1458).
- Entry (j, k) of the matrix in the reversed basis is the coefficient of \mathbf{v}_{n+1-j} in T\mathbf{v}_{n+1-k}, which is entry (n + 1 - j, n + 1 - k) of [T]_{(\mathbf{v}_1, \dots, \mathbf{v}_n)}. Since j < k exactly when n + 1 - j > n + 1 - k, the entries above the diagonal of one matrix are those below the diagonal of the other.
- If n = 1 the only diagonal entry is \lambda. Otherwise run the proof of Result 3 with this \lambda. Since \dim U < n, at least one \mathbf{w}_k is added, and the last one satisfies T\mathbf{w}_r = \lambda\mathbf{w}_r + (\text{a vector of } U), so the last diagonal entry is \lambda.
- A = [T]_{\mathcal{B}} - \lambda I_n is upper triangular, and its diagonal is zero in exactly k places; let J be the other n - k indices. The columns \mathbf{a}_j of A with j \in J are independent: if \sum_{j \in J} c_j\mathbf{a}_j = \mathbf{0} with some c_j \neq 0, let i be the largest index in J with c_i \neq 0. In row i, columns j < i hold 0 and columns j > i carry c_j = 0, so row i of the sum is c_ia_{ii} \neq 0, a contradiction. So \operatorname{rank} A \ge n - k. As A = [T - \lambda I]_{\mathcal{B}}, Result 6 of The matrix of a linear map and the fundamental theorem of linear maps (Null space, range and the fundamental theorem) give \dim E_\lambda = n - \operatorname{rank}(T - \lambda I) \le k.
- By Result 1, S has an eigenvalue \lambda, so E = \ker(S - \lambda I) \neq \{\mathbf{0}\}. If S\mathbf{v} = \lambda\mathbf{v}, then S(T\mathbf{v}) = T(S\mathbf{v}) = \lambda T\mathbf{v}, so E is invariant under T. Result 1 applied to T|_E \in \mathcal{L}(E) gives \mathbf{v} \in E, \mathbf{v} \neq \mathbf{0}, with T\mathbf{v} = \mu\mathbf{v}; also S\mathbf{v} = \lambda\mathbf{v}.
- If T\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} \neq \mathbf{0}, then T^m\mathbf{v} = \lambda^m\mathbf{v} by induction, and T^m = 0 gives \lambda^m = 0, so \lambda = 0. Over \mathbb{C} with V \ne \{\mathbf{0}\}, Result 1 gives T an eigenvalue, which must then be 0. Conversely, if 0 is the only eigenvalue, Example 4 gives T^{\dim V} = 0.
- T(x, y) = (x + y,\ y) on \mathbb{R}^2 has matrix \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix} in (\mathbf{e}_1, \mathbf{e}_2), upper triangular, but T\mathbf{e}_2 = (1, 1) is not a multiple of \mathbf{e}_2.