Invariant subspaces

An operator on a large space is hard to see all at once. Which subspaces does it keep to itself, so that it can be studied one piece at a time?

Where you are 5 in Eigenvalues, proved

Why this exists

An operator T \in \mathcal{L}(V) on a space of dimension 50 is a 50 \times 50 table of numbers, too much to see at once. Suppose instead that V = U_1 \oplus U_2 (Sums and direct sums of subspaces) and that T sends U_1 into U_1 and U_2 into U_2. Then T is two smaller operators side by side, each to be studied alone. If the pieces are lines, T on each is multiplication by a number.

Finding an eigenvector found the eigenspaces of a matrix by row reduction. Here the invariant lines are exactly the eigenvector lines, and kernels of polynomials in T supply many more invariant subspaces. Every later structure theorem about operators, from Upper-triangular matrices over the complex numbers form to the spectral theorem, is built from them.

Definition

An operator on a vector space V over \mathbb{F} is a linear map from V to itself, an element of \mathcal{L}(V).

A subspace U of V is invariant under T \in \mathcal{L}(V) if T\mathbf{u} \in U for every \mathbf{u} \in U. Then the restriction T|_U, defined by T|_U(\mathbf{u}) = T\mathbf{u} for \mathbf{u} \in U, is an operator on U.

A scalar \lambda \in \mathbb{F} is an eigenvalue of T if T\mathbf{v} = \lambda\mathbf{v} for some \mathbf{v} \neq \mathbf{0}, and such a \mathbf{v} is an eigenvector for \lambda. The eigenspace is E_\lambda = \ker(T - \lambda I): the eigenvectors for \lambda together with \mathbf{0}. For a square matrix acting on \mathbb{F}^n these are the definitions of Finding an eigenvector.

For a polynomial p(z) = a_0 + a_1 z + \dots + a_m z^m with coefficients in \mathbb{F}, the operator p(T) is

p(T) = a_0 I + a_1 T + \dots + a_m T^m,

with T^0 = I and T^k the composition of k copies of T.

Intuition

Picture T moving every vector of the plane. A line through the origin is invariant when every vector on it stays on it, so T\mathbf{v} is a multiple of \mathbf{v}: an invariant line is a line of eigenvectors. The figure shows T(x, y) = (y, 2x + y): along \operatorname{span}\{(1, 2)\} every vector doubles, along \operatorname{span}\{(1, -1)\} every vector flips, and the horizontal axis is not invariant, since T(1, 0) = (0, 2).

\mathbf{u}
T\mathbf{u}
\mathbf{v}
T\mathbf{v}
\mathbf{w}
T\mathbf{w}
E_2
E_{-1}
Figure 1The operator T(x, y) = (y, 2x + y) keeps two lines, E_2 and E_{-1}, each mapped into itself. The horizontal axis is not invariant, since T\mathbf{w} leaves it.

In a basis that starts with a basis of an invariant subspace, the matrix of T has a block of zeros (Result 4).

Results

Throughout, V is a vector space over \mathbb{F} and T \in \mathcal{L}(V). Every result holds over any field, with the same proof.

Result 1 (four invariant subspaces). \{\mathbf{0}\}, V, \ker T and \operatorname{im} T are invariant under T.

Result 2 (invariant lines). Let \mathbf{v} \neq \mathbf{0}. The line \operatorname{span}\{\mathbf{v}\} is invariant under T if and only if \mathbf{v} is an eigenvector of T.

Result 3 (tests for an eigenvalue). Let V be finite-dimensional and \lambda \in \mathbb{F}. The following are equivalent: (a) \lambda is an eigenvalue of T; (b) T - \lambda I is not injective; (c) T - \lambda I is not surjective; (d) T - \lambda I is not invertible.

Result 4 (restrictions and block matrices). Let U be invariant under T. Every eigenvalue of T|_U is an eigenvalue of T. If V is finite-dimensional and (\mathbf{u}_1, \dots, \mathbf{u}_k) is a basis of U extended to a basis \mathcal{B} of V, then

[T]_{\mathcal{B}} = \begin{bmatrix} A & X \\ 0 & Y \end{bmatrix}, \qquad A = [T|_U]_{(\mathbf{u}_1, \dots, \mathbf{u}_k)},

where the zero block has k columns. Conversely, if the first k columns of [T]_{\mathcal{B}} are zero below row k, the span of the first k basis vectors is invariant. If V = U \oplus W with W also invariant and \mathcal{B} joins bases of U and W, then X = 0 as well.

Result 5 (polynomials in one operator commute). For polynomials p and q, (pq)(T) = p(T)q(T). Hence p(T)q(T) = q(T)p(T).

Result 6 (kernels and ranges of p(T)). For every polynomial p, the subspaces \ker p(T) and \operatorname{im} p(T) are invariant under T. In particular each eigenspace E_\lambda and each range \operatorname{im}(T - \lambda I) is invariant under T.

Why it is true

Result 1. T\mathbf{0} = \mathbf{0}, and T maps V into V. If \mathbf{u} \in \ker T, then T\mathbf{u} = \mathbf{0}, which lies in the subspace \ker T. If \mathbf{u} \in \operatorname{im} T, then T\mathbf{u} is an output of T, so it lies in \operatorname{im} T. Both are subspaces by Result 1 of Null space, range and the fundamental theorem.

Result 2. If T\mathbf{v} = \lambda\mathbf{v}, then T(c\mathbf{v}) = (c\lambda)\mathbf{v} \in \operatorname{span}\{\mathbf{v}\}. Conversely, if the line is invariant, then T\mathbf{v} \in \operatorname{span}\{\mathbf{v}\}, so T\mathbf{v} = \lambda\mathbf{v} for some \lambda, and \mathbf{v} \neq \mathbf{0} makes \mathbf{v} an eigenvector.

Result 3. T\mathbf{v} = \lambda\mathbf{v} is the same equation as (T - \lambda I)\mathbf{v} = \mathbf{0}, so (a) says \ker(T - \lambda I) \neq \{\mathbf{0}\}, which is (b) by the injectivity test (Result 2 of Null space, range and the fundamental theorem). The operator T - \lambda I maps V to a space of the same finite dimension, so it is injective exactly when it is surjective (Result 6 there): (b) and (c) are equivalent. An operator is invertible exactly when it is injective and surjective (Invertibility and isomorphic spaces), and here those two stand or fall together, so (d) is equivalent to (b).

Result 4. If T|_U\mathbf{u} = \lambda\mathbf{u} with \mathbf{u} \neq \mathbf{0}, the same \mathbf{u} is an eigenvector of T. Column j of [T]_{\mathcal{B}} holds the coordinates of T\mathbf{b}_j (The matrix of a linear map). For j \le k, T\mathbf{u}_j \in U is a combination of \mathbf{u}_1, \dots, \mathbf{u}_k alone, and coordinates are unique (Result 1 of Bases and dimension, proved), so its entries below row k are 0 and the top entries are column j of [T|_U]. Conversely, those zeros say T\mathbf{b}_j \in \operatorname{span}\{\mathbf{b}_1, \dots, \mathbf{b}_k\} for j \le k, and by linearity the image of every vector of that span stays in it. If W is invariant and spanned by the last basis vectors, the same argument gives X = 0.

Result 5. Write p(z) = \sum_j a_j z^j and q(z) = \sum_k b_k z^k, so that (pq)(z) = \sum_{j,k} a_j b_k z^{j+k}. Composition distributes over sums on both sides and lets scalars out, since each T^j is linear, and T^jT^k = T^{j+k} because composition is associative. Hence

\begin{aligned} p(T)q(T) &= \sum_{j,k} a_j b_k T^jT^k \\ &= \sum_{j,k} a_j b_k T^{j+k} = (pq)(T). \end{aligned}

Since pq = qp as polynomials, p(T)q(T) = (pq)(T), which is (qp)(T) = q(T)p(T).

Result 6. Result 5 with q(z) = z gives p(T)T = Tp(T). If p(T)\mathbf{u} = \mathbf{0}, then p(T)(T\mathbf{u}) = T(p(T)\mathbf{u}) = T\mathbf{0} = \mathbf{0}, so T\mathbf{u} \in \ker p(T). If \mathbf{w} = p(T)\mathbf{v}, then T\mathbf{w} = p(T)(T\mathbf{v}) \in \operatorname{im} p(T). Taking p(z) = z - \lambda gives the last sentence.

Worked examples

Example 1

Testing four subspaces of the same space

Let T \in \mathcal{L}(\mathbb{R}^3) send (x_1, x_2, x_3) to (x_1 + 2x_2,\ 3x_2,\ 2x_2 + x_3). Decide which of these are invariant under T, and give the matrix of T|_U for each invariant U: (a) the plane x_2 = 0; (b) \operatorname{span}\{\mathbf{e}_2\}; (c) \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_2\}; (d) \operatorname{span}\{\mathbf{e}_2, (1, 0, 1)\}.

Solution

By linearity, a span is invariant exactly when T sends each spanning vector into it.

(a) T(x_1, 0, x_3) = (x_1, 0, x_3): invariant, and T fixes every vector of it, so in the basis (\mathbf{e}_1, \mathbf{e}_3) the matrix of T|_U is I_2. The plane is E_1.

(b) T\mathbf{e}_2 = (2, 3, 2), not a multiple of \mathbf{e}_2, so by Result 2 the line is not invariant.

(c) T\mathbf{e}_2 leaves the plane x_3 = 0: not invariant.

(d) T\mathbf{e}_2 = (2, 3, 2), and (2, 3, 2) = 3\mathbf{e}_2 + 2(1, 0, 1), while T(1, 0, 1) = (1, 0, 1): invariant. In the basis (\mathbf{e}_2, (1, 0, 1))

[T|_U] = \begin{bmatrix} 3 & 0 \\ 2 & 1 \end{bmatrix}.

Inside this plane lie two eigenvector lines, \operatorname{span}\{(1, 0, 1)\} for 1 and \operatorname{span}\{(1, 1, 1)\} for 3, since (1, 1, 1) = \mathbf{e}_2 + (1, 0, 1) and T(1, 1, 1) = (3, 3, 3).

Example 2

A substitution operator on quadratics

Let T \in \mathcal{L}(\mathcal{P}_2(\mathbb{R})), (Tp)(x) = p(2x + 1). Find every eigenvalue and eigenspace, then \ker and \operatorname{im} of (T - I)(T - 2I).

Solution

T is linear, since substituting into ap + q gives a\,p(2x + 1) + q(2x + 1), and it does not raise degree. If p \neq 0 has degree k and leading coefficient c, then p(2x + 1) has leading term c\,2^kx^k, so Tp = \lambda p forces \lambda = 2^k. The eigenvalues lie among 1, 2, 4. The substitution fixes x = -1, and (2x + 1) + 1 = 2(x + 1), so

T\big((x + 1)^k\big) = 2^k(x + 1)^k, \qquad k = 0, 1, 2.

All three are eigenvalues. If Tp = 2^kp with p \neq 0, then \deg p = k, and r = p - c(x + 1)^k has degree below k with Tr = 2^kr; were r \neq 0, its degree would have to be k. So r = 0 and E_{2^k} = \operatorname{span}\{(x + 1)^k\}.

The list (1, x + 1, (x + 1)^2) has degrees 0, 1, 2, so it is a basis. For p = a + b(x + 1) + c(x + 1)^2, the factor T - 2I multiplies the three terms by -1, 0, 2 and T - I by 0, 1, 3, so

(T - I)(T - 2I)p = 6c(x + 1)^2.

Hence \ker = \operatorname{span}\{1, x + 1\} = \mathcal{P}_1(\mathbb{R}) and \operatorname{im} = \operatorname{span}\{(x + 1)^2\}, both invariant under T, as Result 6 promises.

Example 3

A plane, a block matrix and a third direction

Let A = \begin{bmatrix} 3 & 1 & 0 \\ 1 & 2 & 1 \\ 1 & 1 & 2 \end{bmatrix} act on \mathbb{R}^3 and U = \{\mathbf{x} : x_1 = x_3\}. Show U is invariant, find the matrix of A in the basis \mathcal{B} = ((1, 0, 1), \mathbf{e}_2, \mathbf{e}_3), and use it to find all eigenvalues and eigenvectors.

Solution

The first and third entries of A\mathbf{x} differ by

(3x_1 + x_2) - (x_1 + x_2 + 2x_3) = 2(x_1 - x_3),

so x_1 = x_3 forces equal first and third entries: U is invariant. With \mathbf{u}_1 = (1, 0, 1) and \mathbf{u}_2 = \mathbf{e}_2, a basis of U,

\begin{aligned} A\mathbf{u}_1 &= (3, 2, 3) = 3\mathbf{u}_1 + 2\mathbf{u}_2, \\ A\mathbf{u}_2 &= (1, 2, 1) = \mathbf{u}_1 + 2\mathbf{u}_2, \\ A\mathbf{e}_3 &= (0, 1, 2) = \mathbf{u}_2 + 2\mathbf{e}_3. \end{aligned}

So, as Result 4 predicts,

[A]_{\mathcal{B}} = \begin{bmatrix} 3 & 1 & 0 \\ 2 & 2 & 1 \\ 0 & 0 & 2 \end{bmatrix}.

The block \begin{bmatrix} 3 & 1 \\ 2 & 2 \end{bmatrix} is the matrix of A|_U, with characteristic polynomial \lambda^2 - 5\lambda + 4 = (\lambda - 1)(\lambda - 4) (The characteristic equation and finding eigenvalues). Its eigenvectors have coordinates (1, -2) and (1, 1), that is, \mathbf{u}_1 - 2\mathbf{u}_2 = (1, -2, 1) for 1 and \mathbf{u}_1 + \mathbf{u}_2 = (1, 1, 1) for 4; by Result 4 they are eigenvectors of A.

The corner entry 2 points at a third eigenvalue. A - 2I has equal first and third rows, so it is not invertible, and 2 is an eigenvalue by Result 3, with eigenvector (1, -1, -1). It lies outside U, so its line W meets U only in \mathbf{0}, and \dim U + \dim W = 3 gives \mathbb{R}^3 = U \oplus W (Results 3 and 6 of Sums and direct sums of subspaces). In the basis ((1, -2, 1), (1, 1, 1), (1, -1, -1)) the matrix is \operatorname{diag}(1, 4, 2).

Example 4

No invariant line over the reals

Let T(x, y) = (3x - 5y,\ x - y). Find every subspace of \mathbb{R}^2 invariant under T, then the invariant lines of the same formula on \mathbb{C}^2.

Solution

A subspace of \mathbb{R}^2 has dimension 0, 1 or 2, so besides \{\mathbf{0}\} and \mathbb{R}^2 only lines can be invariant, and by Result 2 an invariant line needs a real eigenvalue. The matrix \begin{bmatrix} 3 & -5 \\ 1 & -1 \end{bmatrix} gives

\det(A - \lambda I) = \lambda^2 - 2\lambda + 2 = (\lambda - 1)^2 + 1,

which is at least 1 for real \lambda. So T has no real eigenvalue, and its only invariant subspaces are \{\mathbf{0}\} and \mathbb{R}^2.

Over \mathbb{C} the roots are 1 \pm i (Complex eigenvalues and rotation-scaling matrices). For 1 + i, the second row of A - (1 + i)I reads z_1 - (2 + i)z_2 = 0, so \mathbf{v} = (2 + i, 1), and A\mathbf{v} = (1 + 3i, 1 + i) = (1 + i)\mathbf{v} because (1 + i)(2 + i) = 1 + 3i. Since A is real, conjugating gives A\bar{\mathbf{v}} = (1 - i)\bar{\mathbf{v}}. The invariant lines on \mathbb{C}^2 are \operatorname{span}\{(2 + i, 1)\} and \operatorname{span}\{(2 - i, 1)\}; neither eigenspace is all of \mathbb{C}^2, since T is not a multiple of I.

Example 5

The rotisserie spit

Yuna and Kofi build a rotisserie with the spit along the x_1-axis. One click of the motor turns the roast by an angle \theta about the spit:

\begin{aligned} R(x_1, x_2, x_3) = (&x_1,\ x_2\cos\theta - x_3\sin\theta, \\ &x_2\sin\theta + x_3\cos\theta). \end{aligned}

Prove that if \sin\theta \neq 0, the only subspaces of \mathbb{R}^3 invariant under R are \{\mathbf{0}\}, the spit L = \operatorname{span}\{\mathbf{e}_1\}, the plane P = \{\mathbf{x} : x_1 = 0\} and \mathbb{R}^3.

Solution

Write c = \cos\theta and s = \sin\theta. R\mathbf{e}_1 = \mathbf{e}_1, and R leaves the first entry alone, so L and P are invariant, as are \{\mathbf{0}\} and \mathbb{R}^3.

Lines. If R\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} = (v_1, v_2, v_3), the last two entries say that \begin{bmatrix} c - \lambda & -s \\ s & c - \lambda \end{bmatrix} sends (v_2, v_3) to \mathbf{0}. That matrix has determinant (c - \lambda)^2 + s^2 > 0, so it is invertible and v_2 = v_3 = 0. Every eigenvector lies on L, and by Result 2 the only invariant line is L.

Planes. Let U be an invariant plane other than P. The map \mathbf{x} \mapsto x_1 from U to \mathbb{R} is linear with null space U \cap P, so by the fundamental theorem of linear maps \dim(U \cap P) \ge 2 - 1 = 1. It is not 2: then U \subseteq P with equal dimensions, so U = P (Result 6 of Bases and dimension, proved). So U \cap P is a line. It is invariant, since \mathbf{x} \in U \cap P gives R\mathbf{x} \in U and R\mathbf{x} \in P. But the only invariant line is L, which meets P only in \mathbf{0}. This contradiction leaves P as the only invariant plane.

Example 6

Every invariant subspace of differentiation

Let D \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), Dp = p'. Prove that the subspaces invariant under D are exactly \{0\}, \mathcal{P}_0(\mathbb{R}), \mathcal{P}_1(\mathbb{R}), \mathcal{P}_2(\mathbb{R}) and \mathcal{P}_3(\mathbb{R}), and deduce that \mathcal{P}_1(\mathbb{R}) has no invariant complement.

Solution

Each \mathcal{P}_k(\mathbb{R}) is \ker D^{k+1}, invariant by Result 6.

Let U \neq \{0\} be invariant and let p \in U have the largest degree k found in U, with leading coefficient c \neq 0. Invariance puts p, Dp, \dots, D^kp in U, and D^jp has degree k - j, with leading coefficient c\,k!/(k - j)! \neq 0. Polynomials of degrees k, k - 1, \dots, 0 are independent: in a vanishing combination, the top-degree one is the only one with an x^k term, so its coefficient is 0, and so on down. They are k + 1 independent vectors in the (k + 1)-dimensional \mathcal{P}_k(\mathbb{R}), so they span it (Result 5 of Bases and dimension, proved), and \mathcal{P}_k(\mathbb{R}) \subseteq U. Every vector of U has degree at most k, so U = \mathcal{P}_k(\mathbb{R}).

A complement of \mathcal{P}_1(\mathbb{R}) has dimension 2, so if invariant it would be \mathcal{P}_1(\mathbb{R}) itself, which is no complement.

Boundaries

  • Over \mathbb{R} an operator may have no eigenvalue and no invariant line (Example 4). Over \mathbb{C}, every operator on a nonzero finite-dimensional space has an eigenvalue; Upper-triangular matrices over the complex numbers proves it.
  • Result 3 needs finite dimension. On \mathcal{P}(\mathbb{R}), multiplication by x (Linear maps between vector spaces) is not surjective, yet it has no eigenvalue at all, since xp has larger degree than p \neq 0.
  • An invariant subspace need not have an invariant complement (Example 6). Then Result 4 gives only a block triangular matrix.
  • Invariance belongs to the pair (U, T). A subspace invariant under T^2 need not be invariant under T (Exercise 18).

Where this goes next

Eigenvectors for distinct eigenvalues are independent shows that an operator on an n-dimensional space has at most n eigenvalues. Upper-triangular matrices over the complex numbers applies Results 4 and 6 to \operatorname{im}(T - \lambda I) over \mathbb{C}. The minimal polynomial studies the p with p(T) = 0, and When an operator is diagonalizable asks when V is a direct sum of eigenspaces.

Exercises

Invariant or not

  1. Let T(x_1, x_2, x_3) = (x_3, x_2, x_1) on \mathbb{R}^3. Which are invariant under T: (a) \operatorname{span}\{(1, 0, 1)\}, (b) \operatorname{span}\{(1, 0, -1)\}, (c) \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_3\}, (d) \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_2\}, (e) \{\mathbf{x} : x_1 + x_3 = 0\}?
  2. For T of Exercise 1, give the matrix of T|_U for U = \{\mathbf{x} : x_1 + x_3 = 0\} in the basis ((1, 0, -1), \mathbf{e}_2), and for U = \operatorname{span}\{\mathbf{e}_1, \mathbf{e}_3\} in the basis (\mathbf{e}_1, \mathbf{e}_3).
  3. Let U = \{p \in \mathcal{P}_3(\mathbb{R}) : p(0) = 0\}. Is U invariant under (a) (Tp)(x) = xp'(x), (b) Dp = p', (c) (Sp)(x) = p(x + 1), (d) (Mp)(x) = p(3x)?
  4. For which k is \operatorname{span}\{(1, k)\} invariant under T(x, y) = (5x + 2y,\ x + 4y)? Is \operatorname{span}\{(0, 1)\}?

Eigenvalues of operators

  1. Find the eigenvalues and eigenspaces of T \in \mathcal{L}(\mathcal{P}_2(\mathbb{R})), (Tp)(x) = p(3 - x).
  2. Find the eigenvalues and eigenspaces of T \in \mathcal{L}(\mathbb{R}^{2 \times 2}), TA = A^{T}, and show \mathbb{R}^{2 \times 2} is the direct sum of them.
  3. Show that T(x, y) = (x - y,\ 5x - 3y) has no invariant line in \mathbb{R}^2, and find its invariant lines in \mathbb{C}^2.
  4. Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), Tp = p + p'. Show that 0 is not an eigenvalue, deduce that T is surjective, find p with Tp = x^3, and find every eigenvalue of T.

Polynomials applied to an operator

  1. For T of Exercise 4, show p(T) = 0 for p(z) = z^2 - 9z + 18, and find \ker and \operatorname{im} of T - 3I and of T - 6I.
  2. Let T(x_1, x_2, x_3) = (x_3, x_1, x_2) and p(z) = 1 + z + z^2. Find \ker p(T) and \operatorname{im} p(T), and show that \ker p(T) contains no line invariant under T.
  3. Let T \in \mathcal{L}(\mathcal{P}_3(\mathbb{R})), (Tp)(x) = (x - 1)p'(x), and p(z) = z^2 - 3z. Find \ker p(T) and \operatorname{im} p(T).

Proofs

  1. Prove that if U and W are invariant under T, so are U + W and U \cap W.
  2. Let S, T \in \mathcal{L}(V) with ST = TS. Prove that \ker S, \operatorname{im} S and every eigenspace of S are invariant under T.
  3. Let V be finite-dimensional, T \in \mathcal{L}(V) invertible and U invariant under T. Prove that U is invariant under T^{-1}.
  4. Prove that if T\mathbf{v} = \lambda\mathbf{v}, then p(T)\mathbf{v} = p(\lambda)\mathbf{v} for every polynomial p. Deduce that if T^2 = I, every eigenvalue of T is 1 or -1.
  5. Let V \neq \{\mathbf{0}\}. Prove that every subspace of V is invariant under T if and only if T = cI for some c \in \mathbb{F}.
  6. Let V = U \oplus W with U and W invariant under T. Prove that \lambda is an eigenvalue of T if and only if it is an eigenvalue of T|_U or of T|_W.
  7. Refute with a counterexample: if U is invariant under T^2, then U is invariant under T.
  8. Kofi sets the motor of Example 5 to half turns, H(x_1, x_2, x_3) = (x_1, -x_2, -x_3). Prove that the subspaces invariant under H are exactly the sums A + B with A equal to \{\mathbf{0}\} or L and B any subspace of P, and list them.
Answers
  1. (a) Yes: T(1, 0, 1) = (1, 0, 1). (b) Yes: T(1, 0, -1) = (-1, 0, 1) = -(1, 0, -1). (c) Yes: T\mathbf{e}_1 = \mathbf{e}_3, T\mathbf{e}_3 = \mathbf{e}_1. (d) No: T\mathbf{e}_1 = \mathbf{e}_3. (e) Yes: T\mathbf{x} = (x_3, x_2, x_1) has x_3 + x_1 = 0 again.
  2. T(1, 0, -1) = -(1, 0, -1) and T\mathbf{e}_2 = \mathbf{e}_2 give \begin{bmatrix} -1 & 0 \\ 0 & 1 \end{bmatrix}; T\mathbf{e}_1 = \mathbf{e}_3 and T\mathbf{e}_3 = \mathbf{e}_1 give \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}.
  3. (a) Yes: (Tp)(0) = 0 \cdot p'(0) = 0. (b) No: Dx = 1. (c) No: Sx = x + 1, which is 1 at 0. (d) Yes: (Mp)(0) = p(0) = 0.
  4. T(1, k) = (5 + 2k, 1 + 4k) is \lambda(1, k) exactly when 1 + 4k = k(5 + 2k), that is 2k^2 + k - 1 = (2k - 1)(k + 1) = 0: k = \tfrac12 (eigenvalue 6) or k = -1 (eigenvalue 3). T(0, 1) = (2, 4), so \operatorname{span}\{(0, 1)\} is not invariant.
  5. With s = 2x - 3, the substitution sends s to 3 - 2x = -s, so T(s^k) = (-1)^ks^k, and (1, s, s^2) is a basis. For p = a + bs + cs^2, Tp = a - bs + cs^2; Tp = \lambda p with p \ne 0 forces \lambda = \pm 1. E_1 = \operatorname{span}\{1, (2x - 3)^2\}, E_{-1} = \operatorname{span}\{2x - 3\}.
  6. T^2 = I, so TA = \lambda A with A \ne 0 gives A = \lambda^2A and \lambda = \pm 1. E_1 is the symmetric matrices, basis \begin{bmatrix} 1 & 0 \\ 0 & 0 \end{bmatrix}, \begin{bmatrix} 0 & 0 \\ 0 & 1 \end{bmatrix}, \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}; E_{-1} = \operatorname{span}\Big\{\begin{bmatrix} 0 & 1 \\ -1 & 0 \end{bmatrix}\Big\}. Every A is \tfrac12(A + A^{T}) + \tfrac12(A - A^{T}), and a matrix in both has A = -A, so A = 0.
  7. \det(A - \lambda I) = \lambda^2 + 2\lambda + 2 = (\lambda + 1)^2 + 1 > 0 for real \lambda, so no real eigenvalue and, by Result 2, no invariant line. Over \mathbb{C}, \lambda = -1 \pm i, with lines \operatorname{span}\{(1, 2 - i)\} for -1 + i and \operatorname{span}\{(1, 2 + i)\} for -1 - i.
  8. Tp = 0 means p' = -p; for p \ne 0 the left side has smaller degree, so p = 0. By Result 3, T is surjective. p = x^3 - 3x^2 + 6x - 6: then p' = 3x^2 - 6x + 6 and p + p' = x^3. Tp = \lambda p means p' = (\lambda - 1)p, which for \lambda \ne 1 forces p = 0 by degree; so 1 is the only eigenvalue, with E_1 = \operatorname{span}\{1\}.
  9. A = \begin{bmatrix} 5 & 2 \\ 1 & 4 \end{bmatrix}, A^2 = \begin{bmatrix} 27 & 18 \\ 9 & 18 \end{bmatrix}, and A^2 - 9A + 18I = 0. A - 3I = \begin{bmatrix} 2 & 2 \\ 1 & 1 \end{bmatrix}: kernel \operatorname{span}\{(1, -1)\}, range \operatorname{span}\{(2, 1)\}. A - 6I = \begin{bmatrix} -1 & 2 \\ 1 & -2 \end{bmatrix}: kernel \operatorname{span}\{(2, 1)\}, range \operatorname{span}\{(1, -1)\}.
  10. p(T)\mathbf{x} = \mathbf{x} + T\mathbf{x} + T^2\mathbf{x} = (s, s, s) with s = x_1 + x_2 + x_3. So \ker p(T) = \{\mathbf{x} : x_1 + x_2 + x_3 = 0\} and \operatorname{im} p(T) = \operatorname{span}\{(1, 1, 1)\}. Since T^3 = I, a real eigenvalue has \lambda^3 = 1, so \lambda = 1, and E_1 = \operatorname{span}\{(1, 1, 1)\}, which meets the plane only in \mathbf{0}. By Result 2 the plane holds no invariant line.
  11. T\big((x - 1)^k\big) = k(x - 1)^k, so p(T) multiplies (x - 1)^k by k^2 - 3k: 0, -2, -2, 0 for k = 0, 1, 2, 3. \ker p(T) = \operatorname{span}\{1, (x - 1)^3\}, \operatorname{im} p(T) = \operatorname{span}\{x - 1, (x - 1)^2\}.
  12. Both are subspaces (Result 1 of Sums and direct sums of subspaces; an intersection of subspaces is closed under the operations). If \mathbf{u} \in U and \mathbf{w} \in W, then T(\mathbf{u} + \mathbf{w}) = T\mathbf{u} + T\mathbf{w} with T\mathbf{u} \in U and T\mathbf{w} \in W, so it lies in U + W. If \mathbf{x} \in U \cap W, then T\mathbf{x} \in U and T\mathbf{x} \in W.
  13. If S\mathbf{v} = \mathbf{0}, then S(T\mathbf{v}) = T(S\mathbf{v}) = \mathbf{0}. If \mathbf{w} = S\mathbf{v}, then T\mathbf{w} = S(T\mathbf{v}) \in \operatorname{im} S. S - \lambda I also commutes with T, since (S - \lambda I)T = ST - \lambda T, which is TS - \lambda T = T(S - \lambda I); so its kernel E_\lambda(S) is invariant under T.
  14. T|_U \in \mathcal{L}(U) is injective, since T is. U is finite-dimensional, so T|_U is surjective (Result 6 of Null space, range and the fundamental theorem). For \mathbf{u} \in U pick \mathbf{u}' \in U with T\mathbf{u}' = \mathbf{u}; then T^{-1}\mathbf{u} = \mathbf{u}' \in U.
  15. By induction T^k\mathbf{v} = \lambda^k\mathbf{v}: true for k = 0, and T^{k+1}\mathbf{v} = T(\lambda^k\mathbf{v}) = \lambda^{k+1}\mathbf{v}. Then p(T)\mathbf{v} = \sum a_kT^k\mathbf{v}, which is \sum a_k\lambda^k\mathbf{v} = p(\lambda)\mathbf{v}. If T^2 = I, take p(z) = z^2 - 1: p(T) = 0, so (\lambda^2 - 1)\mathbf{v} = \mathbf{0} with \mathbf{v} \ne \mathbf{0}, and \lambda = \pm 1.
  16. If T = cI, then T\mathbf{u} = c\mathbf{u} \in U for every subspace U. Conversely, every line is invariant, so by Result 2 every \mathbf{v} \ne \mathbf{0} has T\mathbf{v} = \lambda_{\mathbf{v}}\mathbf{v}. For independent \mathbf{v}, \mathbf{w}, T(\mathbf{v} + \mathbf{w}) is both \lambda_{\mathbf{v}+\mathbf{w}}(\mathbf{v} + \mathbf{w}) and \lambda_{\mathbf{v}}\mathbf{v} + \lambda_{\mathbf{w}}\mathbf{w}, so \lambda_{\mathbf{v}} = \lambda_{\mathbf{v}+\mathbf{w}} = \lambda_{\mathbf{w}}. For \mathbf{w} = a\mathbf{v} with a \ne 0, T\mathbf{w} = a\lambda_{\mathbf{v}}\mathbf{v} = \lambda_{\mathbf{v}}\mathbf{w}. So one c serves every nonzero vector, and T = cI.
  17. An eigenvalue of T|_U or T|_W is one of T (Result 4). Conversely, let T\mathbf{v} = \lambda\mathbf{v}, \mathbf{v} = \mathbf{u} + \mathbf{w} \ne \mathbf{0}. Then T\mathbf{u} + T\mathbf{w} = \lambda\mathbf{u} + \lambda\mathbf{w} with T\mathbf{u}, \lambda\mathbf{u} \in U and T\mathbf{w}, \lambda\mathbf{w} \in W. Decomposition in a direct sum is unique, so T\mathbf{u} = \lambda\mathbf{u} and T\mathbf{w} = \lambda\mathbf{w}. One of \mathbf{u}, \mathbf{w} is nonzero and is an eigenvector of that restriction.
  18. T(x, y) = (-y, x) on \mathbb{R}^2 has T^2 = -I, so every line is invariant under T^2. But T\mathbf{e}_1 = \mathbf{e}_2 \notin \operatorname{span}\{\mathbf{e}_1\}.
  19. Any such sum is invariant: H fixes L pointwise and acts as -I on P, so every subspace of L or P is invariant, and sums of invariant subspaces are (Exercise 12). Conversely, if U is invariant, it is invariant under each H^k, hence under every polynomial in H, in particular \tfrac12(I + H)\mathbf{x} = (x_1, 0, 0) and \tfrac12(I - H)\mathbf{x} = (0, x_2, x_3). Each \mathbf{u} \in U is the sum of these two parts, both in U, so U = (U \cap L) + (U \cap P), and U \cap L is \{\mathbf{0}\} or L. The list: \{\mathbf{0}\}; every line in P; L; every plane containing L; P; \mathbb{R}^3.
math.wiki · CC-BY-SA-4.0