math.wiki
Linear algebra › Linear maps › The matrix of a linear map
Sign in

The matrix of a linear map

A linear map between abstract spaces has no columns to read off. Once a basis is chosen at each end, how does it become a matrix, why does composing maps multiply the matrices, and what happens when the bases change?

Where you are 5 in Linear maps

Why this exists

A linear map T: \mathbb{F}^n \to \mathbb{F}^m is multiplication by its standard matrix (Linear transformation, Result 6 of Linear maps between vector spaces), and composing such maps multiplies their matrices (Composing transformations as a matrix product). Differentiation on \mathcal{P}_3(\mathbb{R}) has no columns, and no basis is forced on it. A basis turns vectors into coordinates (Coordinates in a basis and change of basis); this lesson shows that with a basis at each end every linear map between finite-dimensional spaces is a matrix, that composition becomes the matrix product, and how the matrix changes with the bases.

Mei is an animator, and Jin writes her software. A move is timed by an easing curve, a cubic p on [0, 1]: at the fraction t of the move’s duration, the character has covered the fraction p(t) of the distance. The software stores p by four control values and its speed p' by three; Jin needs the 3 \times 4 matrix between them (Example 1).

Definition

Let V be finite-dimensional over \mathbb{F} with basis \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n). Each \mathbf{v} \in V is c_1\mathbf{v}_1 + \dots + c_n\mathbf{v}_n for exactly one list of scalars (Bases and dimension, proved); its coordinate vector is the column [\mathbf{v}]_{\mathcal{B}} = (c_1, \dots, c_n), and \mathbf{v} \mapsto [\mathbf{v}]_{\mathcal{B}} is the coordinate map.

Let W have basis \mathcal{C} = (\mathbf{w}_1, \dots, \mathbf{w}_m) and T \in \mathcal{L}(V, W). The matrix of T relative to \mathcal{B} and \mathcal{C} is the m \times n matrix whose column k is the coordinate vector of T\mathbf{v}_k:

[T]_{\mathcal{C} \leftarrow \mathcal{B}} = \begin{bmatrix} [T\mathbf{v}_1]_{\mathcal{C}} & \cdots & [T\mathbf{v}_n]_{\mathcal{C}} \end{bmatrix}.

So its entries are the scalars A_{jk} with T\mathbf{v}_k = A_{1k}\mathbf{w}_1 + \dots + A_{mk}\mathbf{w}_m for each k. For an operator with one basis at both ends, [T]_{\mathcal{B}} = [T]_{\mathcal{B} \leftarrow \mathcal{B}}. For two bases of one space, the change-of-coordinates matrix P_{\mathcal{B} \leftarrow \mathcal{B}'} of the written course is [I]_{\mathcal{B} \leftarrow \mathcal{B}'}, the matrix of the identity.

Intuition

The matrix records where the basis goes. By the linear map lemma (Result 5 of Linear maps between vector spaces), T\mathbf{v}_1, \dots, T\mathbf{v}_n determine T, and each is written down by its coordinates. Nothing is lost, but the record depends on the bases.

The coordinate maps translate V and W into \mathbb{F}^n and \mathbb{F}^m, and the matrix is what T becomes:

\begin{array}{ccc} V & \xrightarrow{\quad T \quad} & W \\ \big\downarrow & & \big\downarrow \\ \mathbb{F}^n & \xrightarrow{[T]_{\mathcal{C} \leftarrow \mathcal{B}}} & \mathbb{F}^m \end{array}

Composing maps sets two squares side by side, so matrices multiply; changing bases swaps the downward arrows.

Results

Throughout, U, V, W are finite-dimensional over \mathbb{F}, \dim V = n and \dim W = m. Every result holds, with the same proof, over any field.

Result 1 (the coordinate map). For a basis \mathcal{B} of V, the coordinate map V \to \mathbb{F}^n is linear, injective and surjective.

Result 2 (the matrix computes the map). Let \mathcal{B}, \mathcal{C} be bases of V, W and T \in \mathcal{L}(V, W). For every \mathbf{v} \in V,

[T\mathbf{v}]_{\mathcal{C}} = [T]_{\mathcal{C} \leftarrow \mathcal{B}}\,[\mathbf{v}]_{\mathcal{B}},

and [T]_{\mathcal{C} \leftarrow \mathcal{B}} is the only m \times n matrix with this property.

Result 3 (every matrix is a map). With \mathcal{B}, \mathcal{C} fixed, each A \in \mathbb{F}^{m \times n} is the matrix of exactly one T \in \mathcal{L}(V, W). With every matrix taken relative to \mathcal{B} and \mathcal{C}, [S + T] = [S] + [T] and [aT] = a[T] for S, T \in \mathcal{L}(V, W) and a \in \mathbb{F}.

Result 4 (composition is the product). Let \mathcal{D}, \mathcal{B}, \mathcal{C} be bases of U, V, W, and T \in \mathcal{L}(U, V), S \in \mathcal{L}(V, W). Then

[ST]_{\mathcal{C} \leftarrow \mathcal{D}} = [S]_{\mathcal{C} \leftarrow \mathcal{B}}\,[T]_{\mathcal{B} \leftarrow \mathcal{D}}.

Result 5 (changing bases). Let \mathcal{B}, \mathcal{B}' be bases of V and \mathcal{C}, \mathcal{C}' bases of W. Then P_{\mathcal{B} \leftarrow \mathcal{B}'} is invertible with inverse P_{\mathcal{B}' \leftarrow \mathcal{B}}, and for T \in \mathcal{L}(V, W),

\begin{aligned} &[T]_{\mathcal{C}' \leftarrow \mathcal{B}'} \\ &\quad = P_{\mathcal{C}' \leftarrow \mathcal{C}}\,[T]_{\mathcal{C} \leftarrow \mathcal{B}}\,P_{\mathcal{B} \leftarrow \mathcal{B}'}. \end{aligned}

For an operator and P = P_{\mathcal{B} \leftarrow \mathcal{B}'}, this reads [T]_{\mathcal{B}'} = P^{-1}[T]_{\mathcal{B}}P.

Result 6 (null space and rank). Let A = [T]_{\mathcal{C} \leftarrow \mathcal{B}}. Then \mathbf{v} \in \ker T if and only if [\mathbf{v}]_{\mathcal{B}} \in \ker A, and \mathbf{w} \in \operatorname{im} T if and only if [\mathbf{w}]_{\mathcal{C}} \in \operatorname{Col} A. Hence \dim \ker T = \dim \ker A and \operatorname{rank} T = \operatorname{rank} A.

Why it is true

Result 1. If \mathbf{u} = \sum_k c_k\mathbf{v}_k and \mathbf{v} = \sum_k d_k\mathbf{v}_k, then a\mathbf{u} + \mathbf{v} = \sum_k (ac_k + d_k)\mathbf{v}_k and, coordinates being unique, these are its coordinates. So the map is linear by Result 1 of Linear maps between vector spaces. Coordinates \mathbf{0} belong only to 0\mathbf{v}_1 + \dots + 0\mathbf{v}_n = \mathbf{0}, so the map is injective (Result 2 of Null space, range and the fundamental theorem), and each list (c_1, \dots, c_n) is the coordinate vector of \sum_k c_k\mathbf{v}_k.

Result 2. Let \mathbf{v} = \sum_k c_k\mathbf{v}_k. By Result 1 of Linear maps between vector spaces and Result 1 here,

[T\mathbf{v}]_{\mathcal{C}} = \Big[\sum_k c_kT\mathbf{v}_k\Big]_{\mathcal{C}} = \sum_k c_k[T\mathbf{v}_k]_{\mathcal{C}},

the combination of the columns of the matrix with weights c_k, which is its product with [\mathbf{v}]_{\mathcal{B}} (Multiplying a matrix by a vector). If A' has the property too, apply it to \mathbf{v}_k, whose coordinates are \mathbf{e}_k: column k of A' is A'\mathbf{e}_k = [T\mathbf{v}_k]_{\mathcal{C}}.

Result 3. Let \mathbf{u}_k \in W have column k of A as its coordinates. Since a vector is fixed by its coordinates (Result 1), T has matrix A exactly when T\mathbf{v}_k = \mathbf{u}_k for each k, and by the linear map lemma exactly one T does. Column k of [S + T] is [S\mathbf{v}_k + T\mathbf{v}_k]_{\mathcal{C}}, the sum of the k-th columns by Result 1; scalar multiples are the same.

Result 4. ST is linear (Result 4 of Linear maps between vector spaces). For \mathbf{u} \in U, Result 2 applied twice gives

\begin{aligned} [ST\mathbf{u}]_{\mathcal{C}} &= [S]_{\mathcal{C} \leftarrow \mathcal{B}}\,[T\mathbf{u}]_{\mathcal{B}} \\ &= [S]_{\mathcal{C} \leftarrow \mathcal{B}}\,[T]_{\mathcal{B} \leftarrow \mathcal{D}}\,[\mathbf{u}]_{\mathcal{D}}, \end{aligned}

using associativity of the matrix product (Composing transformations as a matrix product). By the uniqueness in Result 2, the product is [ST]_{\mathcal{C} \leftarrow \mathcal{D}}.

Result 5. Result 4 with both maps I gives P_{\mathcal{B} \leftarrow \mathcal{B}'}P_{\mathcal{B}' \leftarrow \mathcal{B}} = [I]_{\mathcal{B} \leftarrow \mathcal{B}}, which is I_n, and likewise in the other order. Then write T = ITI, with identities of W and V, and apply Result 4 twice along the bases \mathcal{B}', \mathcal{B}, \mathcal{C}, \mathcal{C}'. For an operator, take \mathcal{C} = \mathcal{B} and \mathcal{C}' = \mathcal{B}'.

Result 6. Only \mathbf{0} has coordinates \mathbf{0}, so by Result 2, T\mathbf{v} = \mathbf{0} exactly when A[\mathbf{v}]_{\mathcal{B}} = \mathbf{0}. If \mathbf{w} = T\mathbf{v}, then [\mathbf{w}]_{\mathcal{C}} = A[\mathbf{v}]_{\mathcal{B}} \in \operatorname{Col} A. Conversely, if [\mathbf{w}]_{\mathcal{C}} = A\mathbf{x}, Result 1 gives \mathbf{v} with [\mathbf{v}]_{\mathcal{B}} = \mathbf{x}, and then T\mathbf{v} = \mathbf{w} since their coordinates agree. So the coordinate map of V restricts to an injective linear map from \ker T onto \ker A, and that of W to one from \operatorname{im} T onto \operatorname{Col} A. By the fundamental theorem of linear maps (Result 5 of Null space, range and the fundamental theorem), such a map has domain and range of equal dimension.

Worked examples

Example 1

The speed of an easing curve

Jin’s software stores a cubic by its coordinates in the Bernstein basis \mathcal{B}_3 = (\beta_0, \beta_1, \beta_2, \beta_3) of \mathcal{P}_3(\mathbb{R}) and a quadratic by its coordinates in the basis \mathcal{B}_2 = (\gamma_0, \gamma_1, \gamma_2) of \mathcal{P}_2(\mathbb{R}), where

\begin{aligned} \beta_0 &= (1 - t)^3, & \beta_1 &= 3t(1 - t)^2, \\ \beta_2 &= 3t^2(1 - t), & \beta_3 &= t^3, \\ \gamma_0 &= (1 - t)^2, & \gamma_1 &= 2t(1 - t), \\ \gamma_2 &= t^2. \end{aligned}

These coordinates are the control values. (Example 4 shows both lists are bases.) Find [D]_{\mathcal{B}_2 \leftarrow \mathcal{B}_3} for differentiation D: \mathcal{P}_3(\mathbb{R}) \to \mathcal{P}_2(\mathbb{R}), and the speed of Mei’s curve with control values (0, \tfrac14, 1, 1).

Solution

Differentiating, and using 6t(1 - t) = 3\gamma_1,

\begin{aligned} \beta_0' &= -3\gamma_0, & \beta_1' &= 3\gamma_0 - 3\gamma_1, \\ \beta_2' &= 3\gamma_1 - 3\gamma_2, & \beta_3' &= 3\gamma_2. \end{aligned}

For instance \beta_1' = 3(1 - t)^2 - 6t(1 - t). These coordinates are the columns:

[D]_{\mathcal{B}_2 \leftarrow \mathcal{B}_3} = 3\begin{bmatrix} -1 & 1 & 0 & 0 \\ 0 & -1 & 1 & 0 \\ 0 & 0 & -1 & 1 \end{bmatrix}.

By Result 2, control values (c_0, c_1, c_2, c_3) give a speed with control values 3(c_1 - c_0,\ c_2 - c_1,\ c_3 - c_2), three times the differences of neighbors. For Mei’s curve that is (\tfrac34, \tfrac94, 0): the move starts at speed \tfrac34 and ends at rest. Check: p = \tfrac14\beta_1 + \beta_2 + \beta_3 expands to \tfrac34 t + \tfrac32 t^2 - \tfrac54 t^3, so p' = \tfrac34 + 3t - \tfrac{15}{4}t^2, and \tfrac34\gamma_0 + \tfrac94\gamma_1 expands to the same.

0 1 t 1
p(t)
0 1 t 1 2
p'(t)
Figure 1Mei’s easing curve p, whose coordinates in the Bernstein basis are (0, \tfrac14, 1, 1), drawn with those control values at t = 0, \tfrac13, \tfrac23, 1, and its speed p', with coordinates (\tfrac34, \tfrac94, 0) at t = 0, \tfrac12, 1. Each curve starts at its first control value and ends at its last.
Example 2

Multiplying 2 by 2 matrices on the left

Let M = \begin{bmatrix} 1 & -1 \\ 2 & -2 \end{bmatrix} and T \in \mathcal{L}(\mathbb{R}^{2 \times 2}), TX = MX. Find [T]_{\mathcal{E}} for \mathcal{E} = (E_{11}, E_{12}, E_{21}, E_{22}) and for \mathcal{E}' = (E_{11}, E_{21}, E_{12}, E_{22}), then \ker T, \operatorname{im} T and \operatorname{rank} T.

Solution

T is linear since matrix multiplication distributes and commutes with scalars. ME_{jk} has column j of M as its column k and zeros elsewhere, so TE_{11} = E_{11} + 2E_{21}, TE_{12} = E_{12} + 2E_{22}, TE_{21} = -TE_{11} and TE_{22} = -TE_{12}. As columns,

\begin{aligned} [T]_{\mathcal{E}} &= \begin{bmatrix} 1 & 0 & -1 & 0 \\ 0 & 1 & 0 & -1 \\ 2 & 0 & -2 & 0 \\ 0 & 2 & 0 & -2 \end{bmatrix}, \\ [T]_{\mathcal{E}'} &= \begin{bmatrix} 1 & -1 & 0 & 0 \\ 2 & -2 & 0 & 0 \\ 0 & 0 & 1 & -1 \\ 0 & 0 & 2 & -2 \end{bmatrix}. \end{aligned}

In \mathcal{E}' the coordinates of X list its columns in turn, and T multiplies each column by M. Reordering the basis permuted rows and columns.

[T]_{\mathcal{E}'} has rank 2 and null space given by x_{11} = x_{21}, x_{12} = x_{22}. By Result 6, \operatorname{rank} T = 2, \ker T = \operatorname{span}\{E_{11} + E_{21},\ E_{12} + E_{22}\} (two equal rows), and from the columns \operatorname{im} T = \operatorname{span}\{E_{11} + 2E_{21},\ E_{12} + 2E_{22}\} (second row twice the first).

Example 3

End speeds, by composing

Let E: \mathcal{P}_2(\mathbb{R}) \to \mathbb{R}^2, Eq = (q(0), q(1)), and \mathcal{S} the standard basis of \mathbb{R}^2. Find the matrix of the end-speed map ED relative to \mathcal{B}_3 and \mathcal{S}, and the cubics that start and end at rest.

Solution

At t = 0 only \gamma_0 is nonzero, and at t = 1 only \gamma_2, both with value 1. So [E]_{\mathcal{S} \leftarrow \mathcal{B}_2} has columns (1, 0), (0, 0), (0, 1), and by Result 4 and Example 1,

\begin{aligned} [ED]_{\mathcal{S} \leftarrow \mathcal{B}_3} &= \begin{bmatrix} 1 & 0 & 0 \\ 0 & 0 & 1 \end{bmatrix} [D]_{\mathcal{B}_2 \leftarrow \mathcal{B}_3} \\ &= 3\begin{bmatrix} -1 & 1 & 0 & 0 \\ 0 & 0 & -1 & 1 \end{bmatrix}. \end{aligned}

The end speeds are the slopes of the first and last legs of the control polygon. The matrix has rank 2 and null space given by c_1 = c_0, c_3 = c_2, so by Result 6,

\ker ED = \operatorname{span}\{\beta_0 + \beta_1,\ \beta_2 + \beta_3\}.

The classic ease-in-out curve, \beta_2 + \beta_3 = 3t^2 - 2t^3, starts and ends at rest; Mei’s curve starts at speed \tfrac34.

Example 4

Two bases of the same spaces

Jin’s physics code uses \mathcal{M}_3 = (1, t, t^2, t^3) and \mathcal{M}_2 = (1, t, t^2). Find P_{\mathcal{M}_3 \leftarrow \mathcal{B}_3} and P_{\mathcal{M}_2 \leftarrow \mathcal{B}_2}, show \mathcal{B}_3 and \mathcal{B}_2 are bases, and recover [D]_{\mathcal{M}_2 \leftarrow \mathcal{M}_3} from Example 1 by Result 5.

Solution

Expanding each \beta_k and \gamma_k in powers of t gives the columns

\begin{aligned} P_{\mathcal{M}_3 \leftarrow \mathcal{B}_3} &= \begin{bmatrix} 1 & 0 & 0 & 0 \\ -3 & 3 & 0 & 0 \\ 3 & -6 & 3 & 0 \\ -1 & 3 & -3 & 1 \end{bmatrix}, \\ P_{\mathcal{M}_2 \leftarrow \mathcal{B}_2} &= \begin{bmatrix} 1 & 0 & 0 \\ -2 & 2 & 0 \\ 1 & -2 & 1 \end{bmatrix}. \end{aligned}

Each is triangular with nonzero diagonal, so its columns are independent; by Result 1 so are the \beta_k, and four independent vectors in a space of dimension 4 form a basis (Bases and dimension, proved). Likewise for \mathcal{B}_2.

Going back uses (1 - t) + t = 1: cubing gives 1 = \beta_0 + \beta_1 + \beta_2 + \beta_3, while t\big((1 - t) + t\big)^2 and t^2\big((1 - t) + t\big) give t = \tfrac13\beta_1 + \tfrac23\beta_2 + \beta_3 and t^2 = \tfrac13\beta_2 + \beta_3. So

P_{\mathcal{B}_3 \leftarrow \mathcal{M}_3} = \begin{bmatrix} 1 & 0 & 0 & 0 \\ 1 & \tfrac13 & 0 & 0 \\ 1 & \tfrac23 & \tfrac13 & 0 \\ 1 & 1 & 1 & 1 \end{bmatrix}.

By Result 5 the wanted matrix is P_{\mathcal{M}_2 \leftarrow \mathcal{B}_2}[D]_{\mathcal{B}_2 \leftarrow \mathcal{B}_3}P_{\mathcal{B}_3 \leftarrow \mathcal{M}_3} and the right-hand product is

\begin{bmatrix} 0 & 1 & 0 & 0 \\ 0 & 1 & 1 & 0 \\ 0 & 1 & 2 & 3 \end{bmatrix},

the \mathcal{B}_2-coordinates of D(t^k); for instance 2t = \gamma_1 + 2\gamma_2. Then P_{\mathcal{M}_2 \leftarrow \mathcal{B}_2} on the left gives

[D]_{\mathcal{M}_2 \leftarrow \mathcal{M}_3} = \begin{bmatrix} 0 & 1 & 0 & 0 \\ 0 & 0 & 2 & 0 \\ 0 & 0 & 0 & 3 \end{bmatrix},

the matrix read off directly from D(t^k) = kt^{k-1}.

Example 5

A basis that makes an operator diagonal

Let T \in \mathcal{L}(\mathcal{P}_2(\mathbb{R})), (Tp)(x) = (x + 1)p'(x) - p(x). Find A = [T]_{\mathcal{M}} for \mathcal{M} = (1, x, x^2), then \ker T, \operatorname{im} T and \operatorname{rank} T. Then find [T]_{\mathcal{B}'} for \mathcal{B}' = (1 + x,\ 1,\ (1 + x)^2) and confirm Result 5.

Solution

T is linear: differentiation, then multiplication by x + 1, minus the identity (Results 3 and 4 of Linear maps between vector spaces). T(1) = -1, T(x) = 1 and T(x^2) = 2x(x + 1) - x^2 = 2x + x^2, so

A = \begin{bmatrix} -1 & 1 & 0 \\ 0 & 0 & 2 \\ 0 & 0 & 1 \end{bmatrix}.

Row reduction gives \operatorname{rank} A = 2 and \ker A = \operatorname{span}\{(1, 1, 0)\}. By Result 6, \operatorname{rank} T = 2, \ker T = \operatorname{span}\{1 + x\} and, from the columns, \operatorname{im} T = \operatorname{span}\{1,\ 2x + x^2\}.

In \mathcal{B}': T(1 + x) = 0, T(1) = -1, and T\big((1 + x)^2\big) = 2(1 + x)^2 - (1 + x)^2, which is (1 + x)^2. So [T]_{\mathcal{B}'} = \operatorname{diag}(0, -1, 1). With

P = P_{\mathcal{M} \leftarrow \mathcal{B}'} = \begin{bmatrix} 1 & 1 & 1 \\ 1 & 0 & 2 \\ 0 & 0 & 1 \end{bmatrix},

Result 5 says AP = P\operatorname{diag}(0, -1, 1), and indeed A sends the columns of P to (0, 0, 0), (-1, 0, 0) and (1, 2, 1). The basis comes from u = 1 + x: T sends u^k to ku^k - u^k = (k - 1)u^k.

Example 6

A matrix of zeros and ones

Let T \in \mathcal{L}(V, W) have rank r. Prove that some bases \mathcal{B} of V and \mathcal{C} of W give [T]_{\mathcal{C} \leftarrow \mathcal{B}} with 1 in the first r diagonal places and 0 elsewhere, so that two maps in \mathcal{L}(V, W) of equal rank have equal matrices relative to suitable bases. Find such bases for ED of Example 3.

Solution

Take a basis (\mathbf{u}_1, \dots, \mathbf{u}_k) of \ker T and extend it by \mathbf{v}_1, \dots, \mathbf{v}_r to a basis of V. The proof of the fundamental theorem (Result 5 of Null space, range and the fundamental theorem) shows (T\mathbf{v}_1, \dots, T\mathbf{v}_r) is a basis of \operatorname{im} T, so r is the rank. This independent list extends to a basis \mathcal{C} of W that begins with it. Order V’s basis as \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_r, \mathbf{u}_1, \dots, \mathbf{u}_k) and its column j \le r is [T\mathbf{v}_j]_{\mathcal{C}} = \mathbf{e}_j, while the last k columns are \mathbf{0}. The matrix depends only on r, m and n, so a second map of rank r gets the same one from its own bases.

For ED, r = 2 with \ker ED spanned by \beta_0 + \beta_1 and \beta_2 + \beta_3. Adding \beta_1 and \beta_3 gives four vectors that span, since \beta_0 = (\beta_0 + \beta_1) - \beta_1 and \beta_2 = (\beta_2 + \beta_3) - \beta_3, so they form a basis. From Example 3, ED\beta_1 = (3, 0) and ED\beta_3 = (0, 3). With \mathcal{B} = (\beta_1,\ \beta_3,\ \beta_0 + \beta_1,\ \beta_2 + \beta_3) and \mathcal{C} = ((3, 0), (0, 3)),

[ED]_{\mathcal{C} \leftarrow \mathcal{B}} = \begin{bmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \end{bmatrix}.

Boundaries

  • The matrix depends on the bases. Example 2 gives two matrices for two orders of one basis, and Example 5 a full matrix and a diagonal one for one operator. A property read off a matrix belongs to T only if every choice of bases keeps it: rank does (Result 6), single entries do not, and with free bases at both ends rank is all that survives (Example 6).
  • Finite dimension is needed. D on \mathcal{P}(\mathbb{R}) has no finite basis to be recorded in.
  • Order follows the maps. In [ST] = [S][T] the map applied second stands on the left, and the middle basis of Result 4 must match on both factors.
  • The field fixes the size. Over \mathbb{R}, \mathbb{C} has dimension 2 (Exercise 7).

Where this goes next

Invertibility and isomorphic spaces reads Result 3 as an isomorphism of \mathcal{L}(V, W) with \mathbb{F}^{m \times n}. Dual spaces finds that the dual map has the transpose matrix. Invariant subspaces and Upper-triangular matrices over the complex numbers seek bases giving an operator many zeros, as Exercise 14 begins to, and When an operator is diagonalizable asks when a diagonal one like Example 5’s exists, as Diagonalization did for matrices.

Exercises

Building matrices

  1. Let T: \mathcal{P}_2(\mathbb{R}) \to \mathbb{R}^2, Tp = (p(1), p(2)), and \mathcal{S} the standard basis. Find [T]_{\mathcal{S} \leftarrow \mathcal{B}} for \mathcal{B} = (1, x, x^2) and for \mathcal{B}' = (1,\ x - 1,\ (x - 1)^2).
  2. Let T \in \mathcal{L}(\mathbb{C}^2), T(z, w) = (iz + w,\ z - iw), and \mathcal{B} = ((1, i), (1, -i)). Find [T]_{\mathcal{B}}, and use Results 3 and 4 to show T^2 = 0.
  3. Let T \in \mathcal{L}(\mathbb{R}^{2 \times 2}), TX = X + X^{T}. Find [T]_{\mathcal{E}}, with \mathcal{E} as in Example 2, then \operatorname{rank} T and \ker T.
  4. On \mathcal{P}_2(\mathbb{R}) let (Tp)(x) = p(x + 1) and (Sp)(x) = p(x - 1). Find [T]_{\mathcal{M}} and [S]_{\mathcal{M}} for \mathcal{M} = (1, x, x^2), multiply them in both orders, and say what the products show.
  5. Jin’s software raises a quadratic easing curve to a cubic: \iota: \mathcal{P}_2(\mathbb{R}) \to \mathcal{P}_3(\mathbb{R}), \iota q = q. Find [\iota]_{\mathcal{B}_3 \leftarrow \mathcal{B}_2} by multiplying each \gamma_k by (1 - t) + t, compute [D\iota]_{\mathcal{B}_2} by Result 4, and check its first column directly.
  6. Let R \in \mathcal{L}(\mathcal{P}_1(\mathbb{R})), (Rp)(x) = p(1 - x). Find [R]_{\mathcal{M}} for \mathcal{M} = (1, x) and [R]_{\mathcal{B}'} for \mathcal{B}' = (1,\ 1 - 2x), and confirm Result 5.
  7. View \mathbb{C} as a vector space over \mathbb{R} with basis (1, i). Find the matrices of M_w z = wz for w = 2 - 3i and of conjugation Kz = \bar{z}. Show that [K][M_w] = [M_{\bar{w}}][K] and name the identity it records. Is K linear over \mathbb{C}?

Composition, change of basis and rank

  1. Let T: \mathcal{P}_3(\mathbb{R}) \to \mathbb{R}^{2 \times 2}, Tp = \begin{bmatrix} p(0) & p(1) \\ p(1) & p(2) \end{bmatrix}. Find its matrix relative to (1, x, x^2, x^3) and \mathcal{E}, its rank, and \ker T.
  2. T \in \mathcal{L}(V, W) has [T]_{\mathcal{C} \leftarrow \mathcal{B}} = \begin{bmatrix} 1 & 2 \\ 3 & 4 \end{bmatrix} for \mathcal{B} = (\mathbf{v}_1, \mathbf{v}_2) and \mathcal{C} = (\mathbf{w}_1, \mathbf{w}_2). Find [T]_{\mathcal{C}' \leftarrow \mathcal{B}'} for \mathcal{B}' = (\mathbf{v}_1 + \mathbf{v}_2,\ \mathbf{v}_2) and \mathcal{C}' = (\mathbf{w}_1,\ \mathbf{w}_1 + \mathbf{w}_2), by Result 5 and directly.
  3. For which real a is (Tp)(x) = p(x) + axp'(x) on \mathcal{P}_2(\mathbb{R}) not injective? For each such a, find \ker T.

Proofs

  1. Let S, T \in \mathcal{L}(V) and \mathcal{B} a basis of V. Prove that ST = TS if and only if [S]_{\mathcal{B}}[T]_{\mathcal{B}} = [T]_{\mathcal{B}}[S]_{\mathcal{B}}.
  2. Let T \in \mathcal{L}(V), \mathcal{B} a basis of V and A = [T]_{\mathcal{B}}. Prove that A' = [T]_{\mathcal{B}'} for some basis \mathcal{B}' if and only if A' = P^{-1}AP for some invertible P.
  3. For A \in \mathbb{F}^{m \times n} and B \in \mathbb{F}^{n \times p}, prove \operatorname{rank}(AB) \le \min(\operatorname{rank} A, \operatorname{rank} B) by reading A and B as linear maps.
  4. Let T \in \mathcal{L}(V), \mathcal{B} = (\mathbf{v}_1, \dots, \mathbf{v}_n) a basis and 1 \le k < n. Prove that T maps U = \operatorname{span}\{\mathbf{v}_1, \dots, \mathbf{v}_k\} into U if and only if the entries of [T]_{\mathcal{B}} in rows k + 1, \dots, n of columns 1, \dots, k are all 0.
  5. Let \dim V = \dim W = n and T \in \mathcal{L}(V, W). Prove that [T]_{\mathcal{C} \leftarrow \mathcal{B}} = I_n for some bases \mathcal{B}, \mathcal{C} if and only if T is injective.
Answers
  1. T(1) = (1, 1), T(x) = (1, 2), T(x^2) = (1, 4) give \begin{bmatrix} 1 & 1 & 1 \\ 1 & 2 & 4 \end{bmatrix}. In \mathcal{B}', T(x - 1) = (0, 1) and T((x - 1)^2) = (0, 1) give \begin{bmatrix} 1 & 0 & 0 \\ 1 & 1 & 1 \end{bmatrix}.
  2. \mathcal{B} is a basis, two non-parallel vectors. T(1, i) = (2i, 2) = 2i(1, -i) and T(1, -i) = (0, 0), so [T]_{\mathcal{B}} = \begin{bmatrix} 0 & 0 \\ 2i & 0 \end{bmatrix}. Its square is 0, so [T^2]_{\mathcal{B}} = 0 (Result 4), and only the zero map has that matrix (Result 3).
  3. TE_{11} = 2E_{11}, TE_{12} = TE_{21} = E_{12} + E_{21}, TE_{22} = 2E_{22}, so [T]_{\mathcal{E}} = \begin{bmatrix} 2 & 0 & 0 & 0 \\ 0 & 1 & 1 & 0 \\ 0 & 1 & 1 & 0 \\ 0 & 0 & 0 & 2 \end{bmatrix}, of rank 3, with null space spanned by (0, 1, -1, 0). By Result 6, \ker T = \operatorname{span}\{E_{12} - E_{21}\}, the skew-symmetric matrices.
  4. [T]_{\mathcal{M}} = \begin{bmatrix} 1 & 1 & 1 \\ 0 & 1 & 2 \\ 0 & 0 & 1 \end{bmatrix} from T(x^2) = 1 + 2x + x^2, and [S]_{\mathcal{M}} = \begin{bmatrix} 1 & -1 & 1 \\ 0 & 1 & -2 \\ 0 & 0 & 1 \end{bmatrix}. Both products are I_3, so by Results 4 and 3, ST = TS = I: each shift undoes the other.
  5. \gamma_0 = (1 - t)^3 + t(1 - t)^2 = \beta_0 + \tfrac13\beta_1 and in the same way \gamma_1 = \tfrac23\beta_1 + \tfrac23\beta_2 and \gamma_2 = \tfrac13\beta_2 + \beta_3, so [\iota] = \begin{bmatrix} 1 & 0 & 0 \\ \tfrac13 & \tfrac23 & 0 \\ 0 & \tfrac23 & \tfrac13 \\ 0 & 0 & 1 \end{bmatrix}. Then [D\iota]_{\mathcal{B}_2} = [D][\iota] = \begin{bmatrix} -2 & 2 & 0 \\ -1 & 0 & 1 \\ 0 & -2 & 2 \end{bmatrix}. Directly, \gamma_0' = -2(1 - t) = -2(1 - t)^2 - 2t(1 - t) by (1 - t) + t = 1, which is -2\gamma_0 - \gamma_1.
  6. R(1) = 1 and R(x) = 1 - x, so [R]_{\mathcal{M}} = \begin{bmatrix} 1 & 1 \\ 0 & -1 \end{bmatrix}. R(1 - 2x) = 2x - 1 = -(1 - 2x), so [R]_{\mathcal{B}'} = \operatorname{diag}(1, -1). With P = \begin{bmatrix} 1 & 1 \\ 0 & -2 \end{bmatrix} and P^{-1} = \begin{bmatrix} 1 & \tfrac12 \\ 0 & -\tfrac12 \end{bmatrix}, the product P^{-1}[R]_{\mathcal{M}}P is \operatorname{diag}(1, -1).
  7. M_w(1) = 2 - 3i and M_w(i) = 3 + 2i, so [M_w] = \begin{bmatrix} 2 & 3 \\ -3 & 2 \end{bmatrix}; [K] = \operatorname{diag}(1, -1) and [M_{\bar{w}}] = \begin{bmatrix} 2 & -3 \\ 3 & 2 \end{bmatrix}. Both products are \begin{bmatrix} 2 & 3 \\ 3 & -2 \end{bmatrix}, which by Results 3 and 4 records \overline{wz} = \bar{w}\,\bar{z}. K is not complex-linear: K(i \cdot 1) = -i, while iK(1) = i.
  8. The rows are p(0), p(1), p(1), p(2) in coordinates: \begin{bmatrix} 1 & 0 & 0 & 0 \\ 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 \\ 1 & 2 & 4 & 8 \end{bmatrix}, of rank 3, with null space spanned by (0, 2, -3, 1). So \ker T = \operatorname{span}\{x^3 - 3x^2 + 2x\}, the multiples of x(x - 1)(x - 2).
  9. P_{\mathcal{B} \leftarrow \mathcal{B}'} = \begin{bmatrix} 1 & 0 \\ 1 & 1 \end{bmatrix} and P_{\mathcal{C}' \leftarrow \mathcal{C}} is the inverse of \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix}, namely \begin{bmatrix} 1 & -1 \\ 0 & 1 \end{bmatrix}, so the matrix is \begin{bmatrix} -4 & -2 \\ 7 & 4 \end{bmatrix}. Directly, T(\mathbf{v}_1 + \mathbf{v}_2) = 3\mathbf{w}_1 + 7\mathbf{w}_2, which is -4\mathbf{w}_1 + 7(\mathbf{w}_1 + \mathbf{w}_2), and T\mathbf{v}_2 = 2\mathbf{w}_1 + 4\mathbf{w}_2, which is -2\mathbf{w}_1 + 4(\mathbf{w}_1 + \mathbf{w}_2).
  10. T(x^k) = (1 + ak)x^k, so [T]_{\mathcal{M}} = \operatorname{diag}(1, 1 + a, 1 + 2a), of rank below 3 exactly when a = -1 or a = -\tfrac12 (Result 6). Then \ker T is \operatorname{span}\{x\} or \operatorname{span}\{x^2\} respectively.
  11. By Result 4, [ST]_{\mathcal{B}} = [S]_{\mathcal{B}}[T]_{\mathcal{B}} and [TS]_{\mathcal{B}} = [T]_{\mathcal{B}}[S]_{\mathcal{B}}. By Result 3 each matrix belongs to exactly one operator, so ST = TS exactly when these products are equal.
  12. If A' = [T]_{\mathcal{B}'}, Result 5 gives A' = P^{-1}AP with P = P_{\mathcal{B} \leftarrow \mathcal{B}'} invertible. Conversely, let P be invertible with columns \mathbf{p}_k, and let \mathbf{v}'_k have [\mathbf{v}'_k]_{\mathcal{B}} = \mathbf{p}_k (Result 1). The \mathbf{p}_k are independent (The invertible matrix theorem), so by Result 1 the list \mathcal{B}' = (\mathbf{v}'_1, \dots, \mathbf{v}'_n) is independent, of length \dim V, hence a basis (Bases and dimension, proved). By construction P_{\mathcal{B} \leftarrow \mathcal{B}'} = P, and Result 5 gives [T]_{\mathcal{B}'} = P^{-1}AP = A'.
  13. Let T_A, T_B be multiplication by A and B; in the standard bases their matrices are A and B, so by Results 4 and 6, \operatorname{rank}(AB) = \operatorname{rank} T_AT_B. Every output of T_AT_B is an output of T_A, so this is at most \operatorname{rank} A. Also \operatorname{im} T_AT_B is the range of T_A restricted to \operatorname{im} T_B, of dimension \operatorname{rank} B, so by Result 4 of Null space, range and the fundamental theorem it has dimension at most \operatorname{rank} B.
  14. Column j of [T]_{\mathcal{B}} is [T\mathbf{v}_j]_{\mathcal{B}}, and by uniqueness of coordinates a vector lies in U exactly when its coordinates k + 1, \dots, n are 0. If T maps U into U, then T\mathbf{v}_j \in U for j \le k, so those entries are 0. Conversely, if they are, T\mathbf{v}_j \in U for j \le k, and for \mathbf{u} = \sum_{j \le k} c_j\mathbf{v}_j, T\mathbf{u} = \sum_{j \le k} c_jT\mathbf{v}_j lies in the subspace U.
  15. If [T]_{\mathcal{C} \leftarrow \mathcal{B}} = I_n, Result 6 gives \dim \ker T = \dim \ker I_n = 0, so T is injective (Result 2 of Null space, range and the fundamental theorem). Conversely, if T is injective, the fundamental theorem gives \operatorname{rank} T = n, and Example 6 with r = m = n gives bases in which the matrix is I_n.
math.wiki · CC-BY-SA-4.0