Jordan canonical form

An operator with too few eigenvectors has no diagonal matrix. What is the simplest matrix it can be given, and is that matrix the same whichever basis achieves it?

Where you are 3 in Structure of operators

Why this exists

Diagonalization needs a basis in which the matrix is diagonal, and many operators have none. Generalized eigenvectors goes halfway: over \mathbb{C}, V = G_{\lambda_1} \oplus \dots \oplus G_{\lambda_m}, and on each piece T is \lambda_j I plus a nilpotent operator, but the triangular blocks can still hold anything above the diagonal. How simple can a nilpotent operator’s matrix be made, and is the simplest matrix determined by T itself? The Jordan form answers both: a basis in which the only nonzero entries off the diagonal are 1s directly above it, in blocks whose sizes are fixed by T, so that similarity becomes a count of null space dimensions.

Definition

Throughout, V is a finite-dimensional vector space over \mathbb{F}, n = \dim V and T \in \mathcal{L}(V).

For k \ge 1 and \lambda \in \mathbb{F}, the Jordan block J_k(\lambda) is the k \times k matrix with \lambda on the diagonal, 1 directly above it and 0 elsewhere; J_1(\lambda) = [\lambda]. Write \operatorname{diag}(A_1, \dots, A_p) for the matrix with square blocks A_1, \dots, A_p down the diagonal and zeros elsewhere. A Jordan matrix is one of the form

\operatorname{diag}\big(J_{k_1}(\lambda_1), \dots, J_{k_p}(\lambda_p)\big),

where the \lambda_i need not be distinct. A basis \mathcal{B} is a Jordan basis for T if [T]_{\mathcal{B}} is a Jordan matrix, which is then a Jordan form of T. A matrix A \in \mathbb{F}^{n \times n} has Jordan form J if A = PJP^{-1} with P invertible.

Let N \in \mathcal{L}(V), \mathbf{v} \in V and k \ge 1 with N^k\mathbf{v} = \mathbf{0}. The chain of length k ending at \mathbf{v} is (N^{k-1}\mathbf{v}, \dots, N\mathbf{v}, \mathbf{v}). N sends each entry to the one before it and the first to \mathbf{0}, so the span is invariant, and if the chain is independent, the matrix of N on its span in this basis is J_k(0).

Intuition

Fix an eigenvalue \lambda and let S = T - \lambda I. In a chain for S, S steps each vector down one place and sends the bottom one, an eigenvector, to \mathbf{0}; on the chain’s span T = \lambda I + S has matrix J_k(\lambda). A Jordan basis of G_\lambda is a collection of chains, drawn below as columns of dots.

0
J_{3}(4)
0
J_{2}(4)
0
J_{1}(4)
S
s = 1
3 - 0 = 3
s = 2
5 - 3 = 2
s = 3
6 - 5 = 1
Figure 1The chains of Example 4 for the eigenvalue 4, with S = T - 4I. Each column is a chain and becomes one Jordan block; S moves every dot one row down. Row s holds \dim\ker S^s - \dim\ker S^{s-1} dots, a number fixed by T, so the shape cannot depend on the chains chosen.

The dots in rows 1 to s are a basis of \ker S^s (proof of Result 3), so row s holds \dim\ker S^s - \dim\ker S^{s-1} dots. These numbers belong to T, not to a basis, so the shape, and every block size, is the same whichever chains are chosen. Existence is the harder half: Result 1 builds chains that fit into a basis from the chains of the smaller space \operatorname{im} N.

Results

V is finite-dimensional, n = \dim V and T \in \mathcal{L}(V).

Result 1 (chains for a nilpotent operator). Over any field, let N \in \mathcal{L}(V) be nilpotent. There are vectors \mathbf{v}_1, \dots, \mathbf{v}_p and positive integers k_1, \dots, k_p with N^{k_j}\mathbf{v}_j = \mathbf{0} such that the chains (N^{k_j - 1}\mathbf{v}_j, \dots, \mathbf{v}_j), joined, form a basis of V. In it, [N] = \operatorname{diag}(J_{k_1}(0), \dots, J_{k_p}(0)). (For V = \{\mathbf{0}\}, p = 0.)

Result 2 (Jordan form). Let V be complex, or more generally let m_T be a product of linear factors over \mathbb{F}. Then T has a Jordan basis, and every diagonal entry of a Jordan form of T is an eigenvalue of T.

Result 3 (the blocks are determined by T). Let \operatorname{diag}(J_{k_1}(\mu_1), \dots, J_{k_p}(\mu_p)) be a Jordan form of T, \lambda \in \mathbb{F} and s \ge 1. The number of blocks with \mu_i = \lambda and k_i \ge s is

\dim\ker(T - \lambda I)^s - \dim\ker(T - \lambda I)^{s-1}.

Hence any two Jordan forms of T have the same blocks, counted with repetition, in possibly different orders, and we speak of the Jordan form of T.

Result 4 (reading the Jordan form). Under the hypotheses of Result 2, let \lambda be an eigenvalue of T. In a Jordan form of T: (a) the number of blocks for \lambda is \dim E_\lambda; (b) their sizes add to \dim G_\lambda; (c) the largest has size r_\lambda, the exponent of z - \lambda in m_T. (d) T is diagonalizable if and only if every block is 1 \times 1.

Result 5 (similarity). Two complex n \times n matrices are similar if and only if they have the same Jordan form, up to the order of the blocks.

Result 6 (powers of a Jordan block). Over any field, for k \ge 0 the (s, t) entry of J_m(\lambda)^k is

\binom{k}{t - s}\lambda^{k - t + s} \quad \text{if } 0 \le t - s \le k,

and 0 otherwise, where \lambda^0 = 1 even for \lambda = 0.

Why it is true

Result 1. Induction on n; for n = 0 take p = 0. Let n \ge 1. N is not injective, or every power of N would be, while N^k = 0 on V \neq \{\mathbf{0}\}. So U = \operatorname{im} N has \dim U = n - \dim\ker N < n by the fundamental theorem of linear maps (Null space, range and the fundamental theorem, Result 5). U is invariant and N|_U is nilpotent, so by induction there are \mathbf{u}_1, \dots, \mathbf{u}_p \in U and k_1, \dots, k_p \ge 1 with N^{k_j}\mathbf{u}_j = \mathbf{0} whose chains form a basis of U. Choose \mathbf{w}_j with N\mathbf{w}_j = \mathbf{u}_j.

The vectors N^{k_j - 1}\mathbf{u}_j = N^{k_j}\mathbf{w}_j lie in \ker N and are independent; extend them by \mathbf{z}_1, \dots, \mathbf{z}_q to a basis of \ker N (Result 3 of Bases and dimension, proved), so p + q = \dim\ker N. We claim the chains (N^{k_j}\mathbf{w}_j, \dots, \mathbf{w}_j), of length k_j + 1 since N^{k_j + 1}\mathbf{w}_j = \mathbf{0}, with the one-vector chains (\mathbf{z}_l), form a basis. They number

\begin{aligned} \textstyle\sum_j (k_j + 1) + q &= \dim U + p + q \\ &= \dim\operatorname{im} N + \dim\ker N = n. \end{aligned}

Suppose \sum_j\sum_{i=0}^{k_j} a_{ij}N^i\mathbf{w}_j + \sum_l b_l\mathbf{z}_l = \mathbf{0}. Apply N: as N^{i+1}\mathbf{w}_j = N^i\mathbf{u}_j, zero for i = k_j, and N\mathbf{z}_l = \mathbf{0},

\sum_j\sum_{i=0}^{k_j - 1} a_{ij}N^i\mathbf{u}_j = \mathbf{0}.

These are the basis of U, so a_{ij} = 0 for i < k_j. What remains, \sum_j a_{k_jj}N^{k_j - 1}\mathbf{u}_j + \sum_l b_l\mathbf{z}_l = \mathbf{0}, is a relation in the basis of \ker N, so all coefficients vanish. An independent list of length n is a basis (Result 5 of Bases and dimension, proved). Each chain spans an invariant subspace on which N has matrix J_k(0), so [N] is block diagonal with these blocks.

Result 2. Over \mathbb{C}, m_T splits by the The fundamental theorem of algebra. Let \lambda_1, \dots, \lambda_m be the distinct eigenvalues. By Result 5(a) of Generalized eigenvectors, V = G_{\lambda_1} \oplus \dots \oplus G_{\lambda_m} (for V = \{\mathbf{0}\} the empty basis serves), each G_{\lambda_j} is invariant and N_j = (T - \lambda_jI)|_{G_{\lambda_j}} is nilpotent (Result 3(b) and (d) there). Result 1 gives a basis of G_{\lambda_j} in which [N_j] is a Jordan matrix with zero diagonal, so [T|_{G_{\lambda_j}}] = [N_j] + \lambda_jI is a Jordan matrix with diagonal \lambda_j. Join the bases (Result 5 of Sums and direct sums of subspaces); as each G_{\lambda_j} is invariant, [T] is block diagonal with these blocks. In any Jordan basis, the first vector \mathbf{b} of each block’s part has T\mathbf{b} = \lambda_i\mathbf{b}.

Result 3. Let S = T - \lambda I. In the Jordan basis, [S^s] = \operatorname{diag}(J_{k_1}(\mu_1 - \lambda)^s, \dots, J_{k_p}(\mu_p - \lambda)^s), as block diagonal matrices multiply block by block. A coordinate vector is sent to \mathbf{0} exactly when each of its pieces is, so \dim\ker S^s is the sum over the blocks. If \mu_i \neq \lambda, the block is triangular with only the eigenvalue \mu_i - \lambda \neq 0 (Result 4 of Upper-triangular matrices over the complex numbers), so it and its powers are injective and contribute 0. If \mu_i = \lambda, the block J_k(0)^s sends \mathbf{e}_t to \mathbf{e}_{t-s}, read as \mathbf{0} for t \le s, so its null space is spanned by \mathbf{e}_1, \dots, \mathbf{e}_{\min(s, k)}. So

\dim\ker S^s = \sum_{i:\ \mu_i = \lambda}\min(s, k_i).

As \min(s, k) - \min(s - 1, k) is 1 when k \ge s and 0 otherwise, subtracting the sum for s - 1 counts the blocks for \lambda of size at least s. This count depends only on T, and so does the number of size exactly s, a difference of two counts.

Result 4. (a) Take s = 1. (b) Take s = n: every k_i \le n, so \dim G_\lambda = \sum k_i. (c) \ker S^s \subseteq G_\lambda, with equality exactly when \sum\min(s, k_i) = \sum k_i, that is when s \ge every k_i. The least such s is the largest k_i, and by Result 6(b) of Generalized eigenvectors it is the exponent of z - \lambda in m_T. (d) If every block is 1 \times 1, the Jordan form is diagonal. If T is diagonalizable, a diagonalizing basis is a Jordan basis with 1 \times 1 blocks, and by Result 3 every Jordan form has the same blocks.

Result 5. Let B = P^{-1}AP. If J = Q^{-1}BQ is a Jordan form of B (Result 2), then J = (PQ)^{-1}A(PQ) is one of A, so by Result 3 the two have the same blocks. Conversely, let A = PJP^{-1} and B = QJ'Q^{-1} with J' a reordering of the blocks of J. Reordering the columns of Q chain by chain gives Q' with B = Q'JQ'^{-1}, so B = (Q'P^{-1})A(Q'P^{-1})^{-1}.

Result 6. Let N = J_m(0). As \lambda I and N commute, the binomial theorem gives J_m(\lambda)^k = \sum_{j=0}^{k}\binom{k}{j}\lambda^{k-j}N^j. Since N^j\mathbf{e}_t = \mathbf{e}_{t-j} (or \mathbf{0}), N^j has 1 in each position (s, s + j) and 0 elsewhere, so entry (s, t) receives only the term j = t - s.

Worked examples

Example 1

Reading a Jordan form

An operator T on \mathbb{C}^6 has Jordan form \operatorname{diag}(J_3(2), J_1(2), J_2(-1)). Find p_T, m_T, \dim E_\lambda and \dim G_\lambda for each eigenvalue, and \dim\ker(T - 2I)^s for s = 1, 2, 3, 4. Is T diagonalizable?

Solution

By Result 4, the eigenvalue 2 has two blocks, so \dim E_2 = 2; their sizes give \dim G_2 = 3 + 1 = 4; the largest gives exponent 3 in m_T. The eigenvalue -1 has one block of size 2: \dim E_{-1} = 1, \dim G_{-1} = 2, exponent 2. Hence m_T(z) = (z - 2)^3(z + 1)^2. The Jordan matrix minus \lambda I is triangular, so p_T(\lambda) = (\lambda - 2)^4(\lambda + 1)^2, with sign (-1)^6 = 1. By the display in the proof of Result 3, \dim\ker(T - 2I)^s = \min(s, 3) + \min(s, 1), which is 2, 3, 4, 4. T is not diagonalizable: it has a block larger than 1 \times 1 (Result 4(d)).

Example 2

A Jordan basis in dimension two

Let A = \begin{bmatrix} 3 & -1 \\ 4 & -1 \end{bmatrix}. Find its Jordan form J and an invertible P with P^{-1}AP = J.

Solution

p_A(\lambda) = (3 - \lambda)(-1 - \lambda) + 4 = (\lambda - 1)^2, so 1 is the only eigenvalue and \dim G_1 = 2. A - I_2 = \begin{bmatrix} 2 & -1 \\ 4 & -2 \end{bmatrix} has rank 1, so \dim E_1 = 1: one block (Result 4(a)), and J = J_2(1). Since (A - I_2)^2 = 0, any \mathbf{v} with (A - I_2)\mathbf{v} \neq \mathbf{0} ends a chain of length 2. Take \mathbf{v} = \mathbf{e}_1: (A - I_2)\mathbf{e}_1 = (2, 4), and A(1, 0) = (3, 4) = (2, 4) + (1, 0). So

P = \begin{bmatrix} 2 & 1 \\ 4 & 0 \end{bmatrix}, \qquad P^{-1}AP = \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix}.

Check: AP and PJ both equal \begin{bmatrix} 2 & 3 \\ 4 & 4 \end{bmatrix}. Any \mathbf{v} \notin E_1 would serve, so the Jordan basis is far from unique while J is not.

Example 3

A nilpotent matrix with two chains

Let N = \begin{bmatrix} 1 & 1 & -1 & 0 \\ 1 & 1 & -1 & -1 \\ 1 & 1 & -1 & 0 \\ 1 & 1 & -1 & -1 \end{bmatrix}. Find its Jordan form and a Jordan basis.

Solution

The rows of N are (1, 1, -1, 0) and (1, 1, -1, -1), each twice, so \operatorname{rank} N = 2. The rows of N^2 are (1, 1, -1, -1), \mathbf{0}, (1, 1, -1, -1), \mathbf{0}, so \operatorname{rank} N^2 = 1, and (1, 1, -1, -1) times each column of N is 0, so N^3 = 0. Thus \dim\ker N^s is 2, 3, 4, and by Result 3 there are 2, 1, 1 blocks of size at least 1, 2, 3: the Jordan form is \operatorname{diag}(J_3(0), J_1(0)).

A chain of length 3 needs N^2\mathbf{v} \neq \mathbf{0}. For \mathbf{v} = \mathbf{e}_1, N\mathbf{e}_1 = (1, 1, 1, 1) and N^2\mathbf{e}_1 = (1, 0, 1, 0). The equations of \ker N are x_1 + x_2 - x_3 = 0 and x_4 = 0, with basis (1, 0, 1, 0), (0, 1, 1, 0); the second completes the chain’s eigenvector to a basis of \ker N, as in the proof of Result 1. The Jordan basis is

\big((1, 0, 1, 0), (1, 1, 1, 1), (1, 0, 0, 0), (0, 1, 1, 0)\big),

its matrix P has determinant 1, and P^{-1}NP = \operatorname{diag}(J_3(0), J_1(0)).

Example 4

From null space dimensions to the form

An operator T on \mathbb{C}^8 has only the eigenvalues 4 and -1. For s = 1, 2, 3, 4, \dim\ker(T - 4I)^s is 3, 5, 6, 6, and for s = 1, 2, 3, \dim\ker(T + I)^s is 1, 2, 2. Find the Jordan form of T, m_T and p_T.

Solution

By Result 3, subtracting neighbors in 0, 3, 5, 6, 6 counts the blocks for 4 of size at least s = 1, 2, 3, 4: 3, 2, 1, 0. Subtracting in 0, 1, 2, 2 gives 1, 1, 0 for -1.

For 4 there is one block each of sizes 1, 2 and 3, as in the figure; for -1, one block of size 2. The sizes add to 6 + 2 = 8, matching Result 4(b). The Jordan form is

\operatorname{diag}\big(J_3(4), J_2(4), J_1(4), J_2(-1)\big),

so m_T(z) = (z - 4)^3(z + 1)^2 by Result 4(c), and p_T(\lambda) = (\lambda - 4)^6(\lambda + 1)^2.

Example 5

The shift on polynomials

Let (Tp)(x) = p(x + 1) on \mathcal{P}_3(\mathbb{R}). Find a Jordan basis and the Jordan form of T.

Solution

Let \Delta = T - I, so (\Delta p)(x) = p(x + 1) - p(x). As (x + 1)^k - x^k = kx^{k-1} + \dots, \Delta lowers the degree of a nonconstant p by exactly one and kills constants. So \Delta^4 = 0 and \ker\Delta is the constants. Then (z - 1)^4 annihilates T, so m_T is a power of z - 1 (The minimal polynomial), which splits over \mathbb{R}, and Result 2 applies. As \dim E_1 = 1, there is one block, of size \dim G_1 = 4: the form is J_4(1).

The chain ends at x^3: \Delta x^3 = 3x^2 + 3x + 1, \Delta(3x^2 + 3x + 1) = 3(2x + 1) + 3 = 6x + 6, and \Delta(6x + 6) = 6. In the Jordan basis (6, 6x + 6, 3x^2 + 3x + 1, x^3), [T] = J_4(1). Check one column: T(6x + 6) = 6x + 12 = 6 + (6x + 6).

Example 6

Every complex matrix is similar to its transpose

Prove that every A \in \mathbb{C}^{n \times n} is similar to A^{T}.

Solution

Let R_k be the k \times k matrix with 1s on the antidiagonal, so R_k\mathbf{e}_i = \mathbf{e}_{k+1-i} and R_k^{-1} = R_k. With \mathbf{e}_0 = \mathbf{e}_{k+1} = \mathbf{0}, J_k(\lambda)\mathbf{e}_i = \lambda\mathbf{e}_i + \mathbf{e}_{i-1}, so

\begin{aligned} R_kJ_k(\lambda)R_k\mathbf{e}_i &= R_k(\lambda\mathbf{e}_{k+1-i} + \mathbf{e}_{k-i}) \\ &= \lambda\mathbf{e}_i + \mathbf{e}_{i+1}, \end{aligned}

which is J_k(\lambda)^{T}\mathbf{e}_i. By Result 2, A = PJP^{-1} with J a Jordan matrix. Let R be block diagonal with blocks R_{k_i} matching those of J; then RJR = J^{T} and R^{-1} = R. Transposing,

\begin{aligned} A^{T} &= (P^{T})^{-1}J^{T}P^{T} \\ &= (P^{T})^{-1}RJRP^{T} = Q^{-1}JQ \end{aligned}

with Q = RP^{T}. Since J = P^{-1}AP, A^{T} = (PQ)^{-1}A(PQ).

Example 7

Yuna’s gemstone mines

Yuna plays a phone game about mining gemstones. Each minute every mine yields 2 gemstones, every excavator opens 3 new mines, and excavators are bought only at the start. With \mathbf{x}_k = (g_k, m_k, e_k) the gemstones, mines and excavators after k minutes, \mathbf{x}_{k+1} = A\mathbf{x}_k with A = \begin{bmatrix} 1 & 2 & 0 \\ 0 & 1 & 3 \\ 0 & 0 & 1 \end{bmatrix}. She starts with one excavator and nothing else. Find a Jordan basis for A, a formula for g_k, and g_{60}.

Solution

Let N = A - I_3: N\mathbf{e}_3 = 3\mathbf{e}_2, N\mathbf{e}_2 = 2\mathbf{e}_1, N\mathbf{e}_1 = \mathbf{0}. The chain (6\mathbf{e}_1, 3\mathbf{e}_2, \mathbf{e}_3) is a Jordan basis, in which A has matrix J_3(1): count gemstones in lots of 6 and mines in lots of 3. The start \mathbf{e}_3 has coordinates (0, 0, 1), and by Result 6

J_3(1)^k = \begin{bmatrix} 1 & k & \binom{k}{2} \\ 0 & 1 & k \\ 0 & 0 & 1 \end{bmatrix},

whose third column gives the coordinates of \mathbf{x}_k. So \mathbf{x}_k = \binom{k}{2}6\mathbf{e}_1 + 3k\mathbf{e}_2 + \mathbf{e}_3: g_k = 3k(k - 1) and m_k = 3k. Check: after one minute (0, 3, 1), after two (6, 6, 1). After an hour, g_{60} = 3 \cdot 60 \cdot 59 = 10{,}620 gemstones.

Boundaries

  • m_T must split. The first vector of each block is an eigenvector, so the rotation (x, y) \mapsto (-y, x) of \mathbb{R}^2, with no real eigenvalue, has no Jordan basis over \mathbb{R}; over \mathbb{C} its form is \operatorname{diag}(i, -i).
  • The form jumps under small changes. For \varepsilon \neq 0, \begin{bmatrix} \varepsilon & 1 \\ 0 & 0 \end{bmatrix} has distinct eigenvalues and form \operatorname{diag}(\varepsilon, 0); at \varepsilon = 0 it is J_2(0). So the form cannot be read reliably from approximate entries.
  • Finite dimension is needed. On \mathcal{P}(\mathbb{F}), \ker D is the constants. The first vectors of a basis of finite chains for D would be independent in \ker D, so there would be one chain and \mathcal{P}(\mathbb{F}) would be finite-dimensional.
  • p_T and m_T are not enough from dimension four. \operatorname{diag}(J_2(0), J_2(0)) and \operatorname{diag}(J_2(0), J_1(0), J_1(0)) share both but differ in \dim E_0; in dimension at most three they suffice (Exercise 15).

Where this goes next

The diagonal of a Jordan form lists each eigenvalue \dim G_\lambda times, so Trace and determinant from the structure reads the trace and determinant of a complex operator as the sum and product of its eigenvalues with multiplicity. Result 6 turns powers A^k, as in Example 7, into finite sums of binomial terms.

Exercises

Reading Jordan forms

  1. An operator T on \mathbb{C}^7 has Jordan form \operatorname{diag}(J_2(5), J_2(5), J_1(5), J_2(0)). Find p_T, m_T, \dim E_5, \dim E_0, \operatorname{rank} T and \dim\ker(T - 5I)^s for s = 1, 2, 3.
  2. List, up to block order, every Jordan form of an operator on \mathbb{C}^4 with p_T(\lambda) = (\lambda - 3)^4 and m_T(z) = (z - 3)^2, and say which number tells them apart.
  3. An operator on \mathbb{C}^5 has p_T(\lambda) = -(\lambda - 1)^3(\lambda + 2)^2 and m_T(z) = (z - 1)^2(z + 2). Find its Jordan form.
  4. A nilpotent N on \mathbb{C}^7 has \operatorname{rank} N = 4, \operatorname{rank} N^2 = 2, \operatorname{rank} N^3 = 1 and N^4 = 0. Find its Jordan form.

Jordan bases

  1. Let A = \begin{bmatrix} -2 & 1 \\ -1 & 0 \end{bmatrix}. Find its Jordan form J and an invertible P with P^{-1}AP = J.
  2. Let B = \begin{bmatrix} 2 & 1 & 1 \\ 0 & 2 & 1 \\ 0 & 0 & -1 \end{bmatrix}. Find its Jordan form and a Jordan basis.
  3. Let N = \begin{bmatrix} 1 & -1 & 1 & -1 \\ 1 & -1 & 1 & -1 \\ 0 & 0 & 1 & -1 \\ 0 & 0 & 1 & -1 \end{bmatrix}. Find its Jordan form and a Jordan basis.
  4. Use Result 6 to compute A^{10} for the matrix A of Exercise 5.
  5. Malik plays Yuna’s game with different upgrades: each mine yields 4 gemstones a minute and each excavator opens 1 new mine a minute. He starts with 2 excavators, 5 mines and no gemstones. Find a Jordan basis for his matrix, a formula for g_k, and g_{30}.
  6. Find the Jordan forms of J_3(0)^2 and J_4(0)^2.
  7. Which of A_1 = \begin{bmatrix} 2 & 1 & 0 \\ 0 & 2 & 0 \\ 0 & 0 & 2 \end{bmatrix}, A_2 = \begin{bmatrix} 2 & 0 & 0 \\ 3 & 2 & 0 \\ 0 & 0 & 2 \end{bmatrix} and A_3 = \begin{bmatrix} 2 & 1 & 1 \\ 0 & 2 & 1 \\ 0 & 0 & 2 \end{bmatrix} are similar?

Proofs

  1. Let V be complex and T^m = I for some m \ge 1. Prove that T is diagonalizable.
  2. Let V \neq \{\mathbf{0}\} be complex. Prove that \deg m_T = \deg p_T if and only if \dim E_\lambda = 1 for every eigenvalue \lambda of T.
  3. Let V be complex. Prove that T = D + N with D diagonalizable, N nilpotent and DN = ND.
  4. Prove that two complex n \times n matrices with n \le 3 are similar if and only if they have the same characteristic polynomial and the same minimal polynomial.
  5. Let N be nilpotent on V, n \ge 1, with N^{n-1} \neq 0. Prove that every S \in \mathcal{L}(V) with SN = NS is a polynomial in N.
Answers
  1. p_T(\lambda) = -(\lambda - 5)^5\lambda^2 (sign (-1)^7) and m_T(z) = (z - 5)^2z^2. \dim E_5 = 3, \dim E_0 = 1, and \operatorname{rank} T = 7 - \dim\ker T = 6. \dim\ker(T - 5I)^s = 2\min(s, 2) + \min(s, 1), which is 3, 5, 5.
  2. Sizes adding to \dim G_3 = 4 with largest 2: \operatorname{diag}(J_2(3), J_2(3)), with \dim E_3 = 2, and \operatorname{diag}(J_2(3), J_1(3), J_1(3)), with \dim E_3 = 3. The geometric multiplicity tells them apart.
  3. For 1: sizes adding to 3 with largest 2, so 2, 1. For -2: adding to 2 with largest 1, so 1, 1. The form is \operatorname{diag}(J_2(1), J_1(1), J_1(-2), J_1(-2)).
  4. \dim\ker N^s = 7 - \operatorname{rank} N^s is 3, 5, 6, 7, so the blocks of size at least 1, 2, 3, 4 number 3, 2, 1, 1, and of size exactly 1, 2, 3, 4 number 1, 1, 0, 1: \operatorname{diag}(J_4(0), J_2(0), J_1(0)).
  5. p_A(\lambda) = (\lambda + 1)^2 and A + I_2 = \begin{bmatrix} -1 & 1 \\ -1 & 1 \end{bmatrix} has rank 1, so J = J_2(-1). From \mathbf{e}_1: (A + I_2)\mathbf{e}_1 = (-1, -1), so P = \begin{bmatrix} -1 & 1 \\ -1 & 0 \end{bmatrix}; indeed AP = PJ = \begin{bmatrix} 1 & -2 \\ 1 & -1 \end{bmatrix}.
  6. B is triangular with eigenvalues 2, 2, -1, and B - 2I_3 has rank 2, so \dim E_2 = 1: the form is \operatorname{diag}(J_2(2), J_1(-1)). (B - 2I_3)\mathbf{e}_2 = \mathbf{e}_1 gives the chain (\mathbf{e}_1, \mathbf{e}_2), and (B + I_3)\mathbf{x} = \mathbf{0} gives x_2 = -x_3/3, x_1 = -2x_3/9. Jordan basis \big(\mathbf{e}_1, \mathbf{e}_2, (-2, -3, 9)\big).
  7. N has rank 2 and N^2 = 0, so two blocks, both of size 2: \operatorname{diag}(J_2(0), J_2(0)). N\mathbf{e}_1 = (1, 1, 0, 0) and N\mathbf{e}_3 = (1, 1, 1, 1) are independent, and \big((1, 1, 0, 0), \mathbf{e}_1, (1, 1, 1, 1), \mathbf{e}_3\big) is a Jordan basis (its matrix has determinant 1).
  8. J_2(-1)^{10} = \begin{bmatrix} 1 & -10 \\ 0 & 1 \end{bmatrix} = I_2 - 10J_2(0), so A^{10} = I_2 - 10(A + I_2) = \begin{bmatrix} 11 & -10 \\ 10 & -9 \end{bmatrix}.
  9. A = \begin{bmatrix} 1 & 4 & 0 \\ 0 & 1 & 1 \\ 0 & 0 & 1 \end{bmatrix}, and N = A - I_3 sends \mathbf{e}_3 \mapsto \mathbf{e}_2 \mapsto 4\mathbf{e}_1: Jordan basis (4\mathbf{e}_1, \mathbf{e}_2, \mathbf{e}_3), matrix J_3(1). The start (0, 5, 2) has coordinates (0, 5, 2), sent by J_3(1)^k to \big(5k + 2\binom{k}{2}, 5 + 2k, 2\big), so g_k = 4\big(5k + k(k - 1)\big) = 4k^2 + 16k and g_{30} = 4080.
  10. J_3(0)^2 sends \mathbf{e}_3 \mapsto \mathbf{e}_1 \mapsto \mathbf{0} and \mathbf{e}_2 \mapsto \mathbf{0}: form \operatorname{diag}(J_2(0), J_1(0)). J_4(0)^2 sends \mathbf{e}_4 \mapsto \mathbf{e}_2 \mapsto \mathbf{0} and \mathbf{e}_3 \mapsto \mathbf{e}_1 \mapsto \mathbf{0}: form \operatorname{diag}(J_2(0), J_2(0)).
  11. All have only the eigenvalue 2. A_1 - 2I_3 and A_2 - 2I_3 have rank 1 and square 0, so both have form \operatorname{diag}(J_2(2), J_1(2)); A_3 - 2I_3 has rank 2, one block, J_3(2). So A_1 and A_2 are similar (Result 5), and neither is similar to A_3.
  12. In a Jordan basis (Result 2), J^m = I block by block. A diagonal entry has \lambda^m = 1, so \lambda \neq 0. If a block J_k(\lambda) had k \ge 2, by Result 6 the (1, 2) entry of J_k(\lambda)^m would be m\lambda^{m-1} \neq 0, while I has 0 there. So every block is 1 \times 1, and T is diagonalizable by Result 4(d).
  13. By Result 7(a) of Generalized eigenvectors and Result 4(b), \deg p_T = \sum_\lambda\dim G_\lambda, the sum over eigenvalues of the total size of their blocks; by Result 4(c), \deg m_T = \sum_\lambda r_\lambda, the sum of the largest sizes. For each \lambda the total is at least the largest, with equality exactly when there is one block. So the degrees agree if and only if every eigenvalue has one block, that is \dim E_\lambda = 1 (Result 4(a)).
  14. In a Jordan basis, let D and N be the operators with matrices \operatorname{diag}(\lambda_1I_{k_1}, \dots, \lambda_pI_{k_p}) and \operatorname{diag}(J_{k_1}(0), \dots, J_{k_p}(0)). Then T = D + N and D is diagonalizable. Block by block, (\lambda_iI)J_{k_i}(0) = J_{k_i}(0)(\lambda_iI), so DN = ND, and J_{k_i}(0)^{k_i} = 0 gives N^n = 0.
  15. If B = P^{-1}AP, then \det(B - \lambda I) = \det(P^{-1}(A - \lambda I)P) = \det(A - \lambda I) and q(B) = P^{-1}q(A)P for every polynomial q, so p and m agree. Conversely, by Result 4(b) and (c), for each eigenvalue the multiplicity d in p is the total size of its blocks and the exponent r in m the largest. With d \le 3 the sizes are forced: d = 1 gives 1; d = 2 gives 1, 1 or 2 for r = 1, 2; d = 3 gives 1, 1, 1 or 2, 1 or 3 for r = 1, 2, 3. So p and m fix the Jordan form, and Result 5 gives similarity.
  16. N^n = 0 (Result 3(d) of Generalized eigenvectors, with G_0 = V). Pick \mathbf{v} with N^{n-1}\mathbf{v} \neq \mathbf{0}. If \sum a_iN^i\mathbf{v} = \mathbf{0} with j the least i having a_i \neq 0, applying N^{n-1-j} leaves a_jN^{n-1}\mathbf{v} = \mathbf{0}, which is impossible; so (\mathbf{v}, N\mathbf{v}, \dots, N^{n-1}\mathbf{v}) is independent of length n, a basis. Write S\mathbf{v} = a_0\mathbf{v} + \dots + a_{n-1}N^{n-1}\mathbf{v} = q(N)\mathbf{v}. Then SN^i\mathbf{v} = N^iS\mathbf{v} = q(N)N^i\mathbf{v}, so S and q(N) agree on a basis and S = q(N).
math.wiki · CC-BY-SA-4.0