math.wiki
Linear algebra › Linear maps › Linear maps between vector spaces
Sign in

Linear maps between vector spaces

Differentiation, a delay on a sound recording and multiplication by a matrix all respect adding and scaling. What can be proved about every such map at once, and how do such maps combine?

Where you are 5 in Linear maps

Why this exists

Linear transformation studied functions T: \mathbb{R}^n \to \mathbb{R}^m that respect addition and scaling, each one multiplication by a matrix. Vector spaces from axioms showed that polynomials, sequences and functions obey the same eight rules as columns of numbers. The two ideas meet here: a function between any two vector spaces that respects both operations.

Amara and Kofi run a concert hall’s mixing console. A track is a list of samples (x_1, x_2, x_3, \dots), an element of \mathbb{R}^{\infty}, with zeros after the music stops. A fader multiplies each sample by one number, a delay holds the track back, an echo adds a quieter delayed copy. Each respects mixing and volume: the echo of a mix is the mix of the echoes, and the echo of a louder track is the louder echo. Differentiation and integration behave the same way. Functions like these are the linear maps, and every theorem about them holds for the console, for calculus and for matrices at once.

Definition

Let V and W be vector spaces over the same field \mathbb{F}. A linear map from V to W is a function T: V \to W such that, for all \mathbf{u}, \mathbf{v} \in V and all a \in \mathbb{F},

\begin{aligned} T(\mathbf{u} + \mathbf{v}) &= T\mathbf{u} + T\mathbf{v}, \\ T(a\mathbf{v}) &= a\,T\mathbf{v}. \end{aligned}

The first rule is additivity, the second homogeneity; T\mathbf{v} is short for T(\mathbf{v}). On the left the operations are those of V, on the right those of W. “Linear transformation” means the same thing. \mathcal{L}(V, W) is the set of linear maps from V to W. A linear map from V to itself is an operator on V, and \mathcal{L}(V) = \mathcal{L}(V, V).

For functions S, T: V \to W and a \in \mathbb{F}, the sum and scalar multiple are pointwise: (S + T)\mathbf{v} = S\mathbf{v} + T\mathbf{v} and (aT)\mathbf{v} = a(T\mathbf{v}). For T: U \to V and S: V \to W, the product ST is the composition S \circ T (Combining and composing functions), so (ST)\mathbf{u} = S(T\mathbf{u}): first T, then S. Powers of an operator are T^2 = TT, T^3 = TTT, and so on, with T^0 = I.

The standard examples, each checked below:

  • the zero map 0 \in \mathcal{L}(V, W), 0\mathbf{v} = \mathbf{0}, and the identity operator I \in \mathcal{L}(V), I\mathbf{v} = \mathbf{v};
  • for an m \times n matrix A with entries in \mathbb{F}, the map \mathbf{x} \mapsto A\mathbf{x} from \mathbb{F}^n to \mathbb{F}^m;
  • differentiation D \in \mathcal{L}(\mathcal{P}(\mathbb{R})), Dp = p', linear because (ap + q)' = ap' + q' (Definition of the derivative), and integration f \mapsto \int_a^b f(x)\,dx from C[a, b] to \mathbb{R}, linear by the same rules for the The definite integral;
  • the delay R \in \mathcal{L}(\mathbb{F}^{\infty}), R(x_1, x_2, \dots) = (0, x_1, x_2, \dots), and the advance B(x_1, x_2, \dots) = (x_2, x_3, \dots).

Intuition

A linear map can be applied before or after the vector-space operations with the same result, so it sends a combination to the same combination of the images (Result 1). Two consequences shape the subject: a linear map is fixed by its values on a basis (Result 5), which is why a matrix can record it; and linear maps can be added, scaled and composed with linear results (Results 3 and 4).

On a single sample, a console effect is a function from \mathbb{R} to \mathbb{R}; the linear ones are multiplication by a fixed number (Result 2), with graphs that are lines through the origin.

fader
y = 0.6x
offset
y = 0.6x + 0.4
clipper
y = \max(-1, \min(1, x))
Figure 1Three console effects acting on a single sample. Only the fader, x \mapsto 0.6x, is linear, and its graph is a line through the origin. The offset drawn here, x \mapsto 0.6x + 0.4, misses the origin, and the clipper passes through the origin but bends.

Results

Throughout, U, V, W are vector spaces over \mathbb{F}. Every result holds, with the same proof, over any field.

Result 1 (linear maps preserve combinations). Let T: V \to W be linear. Then T\mathbf{0} = \mathbf{0}, T(-\mathbf{v}) = -T\mathbf{v}, and for all \mathbf{v}_1, \dots, \mathbf{v}_k \in V and a_1, \dots, a_k \in \mathbb{F},

\begin{aligned} &T(a_1\mathbf{v}_1 + \dots + a_k\mathbf{v}_k) \\ &\quad = a_1T\mathbf{v}_1 + \dots + a_kT\mathbf{v}_k. \end{aligned}

Conversely, a function T: V \to W is linear if and only if T(a\mathbf{u} + \mathbf{v}) = aT\mathbf{u} + T\mathbf{v} for all \mathbf{u}, \mathbf{v} \in V and a \in \mathbb{F}.

Result 2 (maps out of the scalars). For each \mathbf{w} \in W the map a \mapsto a\mathbf{w} from \mathbb{F} to W is linear, and every linear T: \mathbb{F} \to W has this form, with \mathbf{w} = T(1). In particular every operator on \mathbb{F} is multiplication by the scalar T(1).

Result 3 (the space of linear maps). With the pointwise sum and scalar multiple, \mathcal{L}(V, W) is a vector space over \mathbb{F}, a subspace of the space of all functions from V to W. Its zero vector is the zero map.

Result 4 (products). If T \in \mathcal{L}(U, V) and S \in \mathcal{L}(V, W), then ST \in \mathcal{L}(U, W). Whenever the products are defined, with I the identity on the appropriate space,

  • (T_1T_2)T_3 = T_1(T_2T_3);
  • IT = TI = T;
  • (S_1 + S_2)T = S_1T + S_2T and S(T_1 + T_2) = ST_1 + ST_2;
  • (aS)T = a(ST) = S(aT) for a \in \mathbb{F}.

The product need not commute: ST \neq TS is possible even when both are defined (Example 4).

Result 5 (the linear map lemma). Let (\mathbf{b}_1, \dots, \mathbf{b}_n) be a basis of V, a list such that each vector of V is a combination of it in exactly one way (Bases and dimension, proved), and let \mathbf{w}_1, \dots, \mathbf{w}_n \in W be any vectors, repeats allowed. Exactly one T \in \mathcal{L}(V, W) has T\mathbf{b}_k = \mathbf{w}_k for k = 1, \dots, n, namely

\begin{aligned} &T(c_1\mathbf{b}_1 + \dots + c_n\mathbf{b}_n) \\ &\quad = c_1\mathbf{w}_1 + \dots + c_n\mathbf{w}_n. \end{aligned}

Result 6 (maps between lists of scalars). A function T: \mathbb{F}^n \to \mathbb{F}^m is linear if and only if T\mathbf{x} = A\mathbf{x} for some m \times n matrix A, and then A = [\,T\mathbf{e}_1 \ \cdots \ T\mathbf{e}_n\,] is the only such matrix. Linear transformation proved this over \mathbb{R}.

Why it is true

Result 1. By additivity, T\mathbf{0} = T(\mathbf{0} + \mathbf{0}), which is T\mathbf{0} + T\mathbf{0}. Since also T\mathbf{0} = T\mathbf{0} + \mathbf{0}, cancellation in W (Result 3 of Vector spaces from axioms) gives T\mathbf{0} = \mathbf{0}. By homogeneity and Result 6 of that lesson, T(-\mathbf{v}) = T((-1)\mathbf{v}) = (-1)T\mathbf{v} = -T\mathbf{v}. The combination rule is homogeneity when k = 1. If it holds for k terms, additivity splits T of a sum of k + 1 terms into T of the first k plus T(a_{k+1}\mathbf{v}_{k+1}), and the hypothesis and homogeneity finish the step (Proof by induction).

For the criterion: a linear T splits T(a\mathbf{u} + \mathbf{v}) into T(a\mathbf{u}) + T\mathbf{v} by additivity, and then into aT\mathbf{u} + T\mathbf{v} by homogeneity. Conversely, suppose the identity holds. With a = 1 it is additivity. With a = 1 and \mathbf{u} = \mathbf{v} = \mathbf{0} it gives T\mathbf{0} = T\mathbf{0} + T\mathbf{0}, so T\mathbf{0} = \mathbf{0}, and then \mathbf{v} = \mathbf{0} gives T(a\mathbf{u}) = aT\mathbf{u}.

Result 2. Additivity and homogeneity of a \mapsto a\mathbf{w} say (a + b)\mathbf{w} = a\mathbf{w} + b\mathbf{w} and (ca)\mathbf{w} = c(a\mathbf{w}), which are Axioms 8 and 6. If T: \mathbb{F} \to W is linear, then a = a \cdot 1 in \mathbb{F}, and homogeneity gives T(a) = T(a \cdot 1) = a\,T(1).

Result 3. The functions from V to W with pointwise operations form a vector space: the proof of Result 8 of Vector spaces from axioms goes through word for word with values in W, each axiom checked at a point by the same axiom in W. By the subspace test (Result 9 there), it is enough that the zero map is linear, which holds since \mathbf{0} + \mathbf{0} = \mathbf{0} and a\mathbf{0} = \mathbf{0}, and that sums and multiples of linear maps are linear. For linear S and T, scalars a, b and vectors \mathbf{u}, \mathbf{v}, Axioms 1, 2 and 7 give

\begin{aligned} &(S + T)(a\mathbf{u} + \mathbf{v}) \\ &\quad = aS\mathbf{u} + S\mathbf{v} + aT\mathbf{u} + T\mathbf{v} \\ &\quad = a(S + T)\mathbf{u} + (S + T)\mathbf{v}, \end{aligned}

and Axioms 7 and 6 with ba = ab in \mathbb{F} turn (bT)(a\mathbf{u} + \mathbf{v}) into a(bT)\mathbf{u} + (bT)\mathbf{v}. Result 1’s criterion finishes both.

Result 4. For linear S and T, (ST)(a\mathbf{u} + \mathbf{v}) is S(aT\mathbf{u} + T\mathbf{v}), which is aST\mathbf{u} + ST\mathbf{v} by linearity of S, so ST is linear by Result 1. Maps are equal when they agree at every vector \mathbf{u}. Both sides of the first identity send \mathbf{u} to T_1(T_2(T_3\mathbf{u})), and IT, TI both send it to T\mathbf{u}. ((S_1 + S_2)T)\mathbf{u} = S_1(T\mathbf{u}) + S_2(T\mathbf{u}) by the definition of the sum, while (S(T_1 + T_2))\mathbf{u} = S(T_1\mathbf{u} + T_2\mathbf{u}) = ST_1\mathbf{u} + ST_2\mathbf{u} needs additivity of S. Finally ((aS)T)\mathbf{u} = a\,S(T\mathbf{u}), and (S(aT))\mathbf{u} = S(aT\mathbf{u}) = a\,S(T\mathbf{u}) by homogeneity of S.

Result 5. Existence. Each \mathbf{v} \in V is c_1\mathbf{b}_1 + \dots + c_n\mathbf{b}_n for exactly one list of scalars, so the displayed formula assigns one vector to each \mathbf{v}: it defines a function. Taking c_k = 1 and every other coefficient 0 gives T\mathbf{b}_k = \mathbf{w}_k. If \mathbf{u} = \sum d_k\mathbf{b}_k and \mathbf{v} = \sum c_k\mathbf{b}_k, the axioms give a\mathbf{u} + \mathbf{v} = \sum (ad_k + c_k)\mathbf{b}_k, and by uniqueness these are its coefficients. So T(a\mathbf{u} + \mathbf{v}) = \sum (ad_k + c_k)\mathbf{w}_k, which is aT\mathbf{u} + T\mathbf{v}, and T is linear by Result 1. Uniqueness. If T' is linear with T'\mathbf{b}_k = \mathbf{w}_k, Result 1 gives T'(\sum c_k\mathbf{b}_k) = \sum c_k\mathbf{w}_k, so T' and T agree at every vector.

Result 6. The map \mathbf{x} \mapsto A\mathbf{x} is linear: A(a\mathbf{x} + \mathbf{y}) = aA\mathbf{x} + A\mathbf{y} by the arithmetic of Multiplying a matrix by a vector, which uses only the field laws. If T is linear and A has columns T\mathbf{e}_1, \dots, T\mathbf{e}_n, then \mathbf{x} \mapsto A\mathbf{x} and T agree on the standard basis, so they are equal by the uniqueness in Result 5. If B also works, column j of B is B\mathbf{e}_j = T\mathbf{e}_j, so B = A.

Worked examples

Example 1

An echo on the console

Kofi’s echo adds a copy of the track delayed by three samples at half volume: E = I + \tfrac12 R^3, where R is the delay. Show that R is linear, conclude that E is linear, and find E\mathbf{x} for \mathbf{x} = (4, -2, 0, 2, 0, 0, \dots).

Solution

Sequences are added and scaled sample by sample, so

\begin{aligned} &R(a\mathbf{x} + \mathbf{y}) \\ &\quad = (0, ax_1 + y_1, ax_2 + y_2, \dots) \\ &\quad = a(0, x_1, x_2, \dots) \\ &\qquad + (0, y_1, y_2, \dots), \end{aligned}

which is aR\mathbf{x} + R\mathbf{y}. By Result 1, R is linear. By Result 4, R^3 = RRR is linear, and by Result 3 so is I + \tfrac12 R^3: the echo needs no check of its own.

Delaying three times gives R^3\mathbf{x} = (0, 0, 0, 4, -2, 0, 2, 0, \dots) and

\begin{aligned} E\mathbf{x} &= (4, -2, 0, 2, 0, 0, 0, 0, \dots) \\ &\quad + (0, 0, 0, 2, -1, 0, 1, 0, \dots) \\ &= (4, -2, 0, 4, -1, 0, 1, 0, \dots). \end{aligned}
Example 2

Three more effects

Amara tests three more effects on tracks in \mathbb{R}^{\infty}: an offset N that adds 0.1 to every sample, a clipper C that replaces each sample above 1 by 1 and each sample below -1 by -1, and a smoother G whose sample k is \tfrac12(x_k + x_{k+1}). Which are linear?

Solution

The offset. N\mathbf{0} = (0.1, 0.1, 0.1, \dots) \neq \mathbf{0}, so by Result 1, N is not linear.

The clipper. C\mathbf{0} = \mathbf{0} and C(-\mathbf{x}) = -C\mathbf{x}, which proves nothing. Homogeneity fails: for \mathbf{x} = (0.8, 0, 0, \dots), C(2\mathbf{x}) = (1, 0, 0, \dots), while 2C\mathbf{x} = (1.6, 0, 0, \dots). Not linear.

The smoother. Sample k of G\mathbf{x} is \tfrac12 x_k + \tfrac12 x_{k+1}, so G = \tfrac12(I + B) with B the advance. B is linear by the computation used for R, so G is linear by Result 3.

Example 3

A map from polynomials to triples

Define T: \mathcal{P}_2(\mathbb{R}) \to \mathbb{R}^3 by Tp = \big(p(0),\ p'(1),\ \int_{-1}^{1} p(x)\,dx\big). Show that T is linear, find T(3 - x + 2x^2), and recompute it from T(1), T(x) and T(x^2).

Solution

For p, q \in \mathcal{P}_2(\mathbb{R}) and a \in \mathbb{R}: (ap + q)(0) = ap(0) + q(0), as polynomials are added and scaled pointwise; (ap + q)' = ap' + q', so the same holds at 1; and the integral is linear. So each entry of T(ap + q) is a times the entry of Tp plus that of Tq, that is, T(ap + q) = aTp + Tq, and T is linear by Result 1.

For p = 3 - x + 2x^2: p(0) = 3; p'(x) = -1 + 4x, so p'(1) = 3; and \int_{-1}^{1}(3 - x + 2x^2)\,dx = 6 - 0 + \tfrac43. So Tp = (3, 3, \tfrac{22}{3}).

On the basis (1, x, x^2): T(1) = (1, 0, 2), T(x) = (0, 1, 0) and T(x^2) = (0, 2, \tfrac23). By Result 1,

\begin{aligned} Tp &= 3T(1) - T(x) + 2T(x^2) \\ &= (3, 0, 6) - (0, 1, 0) + (0, 4, \tfrac43) \\ &= (3, 3, \tfrac{22}{3}). \end{aligned}

Three vectors carry all of T (Result 5).

Example 4

Differentiate, then multiply by x

On \mathcal{P}(\mathbb{R}) let D be differentiation, M multiplication by x, (Mp)(x) = xp(x), and \tau the translation (\tau p)(x) = p(x + 1). Show that DM - MD = I and D\tau = \tau D, and check the first on p = 5 - x^2.

Solution

M is linear, since x(ap(x) + q(x)) equals a\,xp(x) + xq(x), and \tau likewise. By the product rule, (DMp)(x) = (xp(x))', which is p(x) + xp'(x), while (MDp)(x) = xp'(x). So (DM - MD)p = p for every p, that is, DM - MD = I. In particular DM \neq MD. By the chain rule, (D\tau p)(x) = \tfrac{d}{dx}\,p(x + 1), which is p'(x + 1) = (\tau Dp)(x), so D and \tau do commute.

For p = 5 - x^2: Mp = 5x - x^3 and DMp = 5 - 3x^2; Dp = -2x and MDp = -2x^2. The difference is 5 - 3x^2 + 2x^2 = 5 - x^2 = p.

Example 5

A map fixed by its values on a basis

Find the linear T: \mathcal{P}_2(\mathbb{R}) \to \mathbb{R}^2 with T(1) = (2, 0), T(1 + x) = (1, 3) and T(1 + x + x^2) = (0, -1). Give T(a_0 + a_1x + a_2x^2) and T(4 + x - 2x^2).

Solution

First, (1,\ 1 + x,\ 1 + x + x^2) is a basis. Writing a_0 + a_1x + a_2x^2 = c_1 \cdot 1 + c_2(1 + x) + c_3(1 + x + x^2) and comparing coefficients of x^2, x and 1 gives c_3 = a_2, c_2 + c_3 = a_1 and c_1 + c_2 + c_3 = a_0, whose only solution is c_1 = a_0 - a_1, c_2 = a_1 - a_2, c_3 = a_2. So each polynomial is a combination of the list in exactly one way, and Result 5 gives exactly one such T:

\begin{aligned} &T(a_0 + a_1x + a_2x^2) \\ &\ = (a_0 - a_1)(2, 0) + (a_1 - a_2)(1, 3) \\ &\qquad + a_2(0, -1) \\ &\ = (2a_0 - a_1 - a_2,\ 3a_1 - 4a_2). \end{aligned}

For 4 + x - 2x^2 the formula gives (8 - 1 + 2,\ 3 + 8) = (9, 11). Through the coordinates instead: c_1 = 3, c_2 = 3, c_3 = -2, and 3(2, 0) + 3(1, 3) - 2(0, -1) = (9, 11).

Example 6

Keeping the opening of a track

Let K \in \mathcal{L}(V) satisfy K^2 = K. Prove that (I - K)^2 = I - K and K(I - K) = (I - K)K = 0. Then read this for Amara’s trim K(x_1, x_2, \dots) = (x_1, x_2, x_3, 0, 0, \dots) on \mathbb{R}^{\infty}.

Solution

Here I - K means I + (-1)K. By the distributive laws and IK = KI = K (Result 4),

\begin{aligned} (I - K)^2 &= (I - K)I - (I - K)K \\ &= I - K - K + K^2 \\ &= I - K, \end{aligned}

where the last step replaces K^2 by K. In the same way, K(I - K) = K - K^2 = 0 and (I - K)K = K - K^2 = 0.

The trim is linear (each kept sample of a\mathbf{x} + \mathbf{y} is ax_k + y_k), and trimming twice is trimming once, so K^2 = K. Then (I - K)\mathbf{x} = (0, 0, 0, x_4, x_5, \dots) is the rest of the track, \mathbf{x} = K\mathbf{x} + (I - K)\mathbf{x}, and K(I - K) = 0 says the rest has a silent opening.

Example 7

Real-linear but not complex-linear

Let T(z, w) = (\bar{w}, \bar{z}) on \mathbb{C}^2. Show that T is additive and that T(r\mathbf{v}) = rT\mathbf{v} for real r, but that T is not a linear map of the complex vector space \mathbb{C}^2.

Solution

The conjugate of a sum is the sum of the conjugates (Complex numbers), so T is additive. A real r has \overline{rw} = r\bar{w}, so T(r\mathbf{v}) = rT\mathbf{v}. Thus T is linear on \mathbb{C}^2 with real scalars, as in Example 6 of Vector spaces from axioms.

With complex scalars homogeneity fails. For \mathbf{v} = (1, 0), T(i\mathbf{v}) = T(i, 0) = (0, -i), but iT\mathbf{v} = i(0, 1) = (0, i). In general T(c\mathbf{v}) = \bar{c}\,T\mathbf{v}. Whether a map is linear depends on the scalars.

Boundaries

  • Both rules, and the field, matter. Additivity alone gives homogeneity for rational scalars (Exercise 19) but not for i (Example 7). Homogeneity alone is not enough either: Linear transformation shows \displaystyle \sqrt[3]{x^3 + y^3} is homogeneous and not additive.
  • An offset is never linear. A map \mathbf{v} \mapsto T\mathbf{v} + \mathbf{w}_0 with T linear and \mathbf{w}_0 \neq \mathbf{0} sends \mathbf{0} to \mathbf{w}_0 (Example 2).
  • DM - MD = I needs infinite dimension. For n \times n matrices, \operatorname{tr}(AB) = \sum_{i,j} a_{ij}b_{ji} = \operatorname{tr}(BA), so AB - BA has trace 0, while I_n has trace n \neq 0 in \mathbb{R} or \mathbb{C} (not modulo a prime p = n). By Result 6, no operators on \mathbb{F}^n have ST - TS = I; no finite list spans \mathcal{P}(\mathbb{R}).
  • A one-sided inverse need not be two-sided. On \mathbb{F}^{\infty}, BR = I but RB \neq I (Exercise 16): the delay is injective (one-to-one) but not surjective (onto) (One-to-one, onto and bijective functions). In finite dimension one product equal to I forces the other (Invertibility and isomorphic spaces).

Where this goes next

Null space, range and the fundamental theorem studies the vectors a linear map sends to \mathbf{0} and the vectors it reaches. The matrix of a linear map uses Result 5 to record T as a matrix [T]_{\mathcal{C} \leftarrow \mathcal{B}} and turns ST into a matrix product, as Composing transformations as a matrix product did in \mathbb{R}^n. Invertibility and isomorphic spaces asks when a linear map can be undone, Dual spaces studies \mathcal{L}(V, \mathbb{F}), Invariant subspaces studies a single operator, and Adjoints and self-adjoint operators moves an operator across an inner product.

Exercises

Deciding linearity

  1. Show that T: \mathcal{P}_2(\mathbb{R}) \to \mathbb{R}, Tp = p(2) - p(-1), is linear, and find T(1 + 3x - x^2).
  2. Show that T: \mathbb{R}^{2 \times 2} \to \mathbb{R}^{2 \times 2}, TA = A - A^{T}, is linear, and find T\begin{bmatrix} 1 & 4 \\ 2 & 3 \end{bmatrix}.
  3. Is A \mapsto \det A a linear map from \mathbb{R}^{2 \times 2} to \mathbb{R}?
  4. Is the map on \mathbb{R}^{\infty} whose sample k is x_kx_{k+1} linear?
  5. Show that Tf = \int_0^1 xf(x)\,dx is a linear map from C[0, 1] to \mathbb{R}, and find T of the function e^x.
  6. Amara’s rectifier replaces each sample x_k of a track by \lvert x_k \rvert. Is it linear?

Computing with linear maps

  1. With R the delay, find (I - \tfrac13 R^2)\mathbf{x} for \mathbf{x} = (3, 0, 6, 0, 0, \dots).
  2. With D and M as in Example 4 and p = 1 + 2x - x^2, find DMp, MDp and D^2Mp, and confirm that DMp - MDp = p.
  3. For S(x, y) = (y, x) and T(x, y) = (x + y, 0) on \mathbb{R}^2, find formulas for ST, TS and S + 2T.
  4. Find the linear T: \mathbb{R}^2 \to \mathbb{R}^3 with T(1, 2) = (1, 0, 1) and T(1, 3) = (0, 1, 1). Give T(x, y) and T(5, 4).
  5. For which c is there a linear T: \mathbb{R}^2 \to \mathbb{R}^2 with T(1, 1) = (1, 2), T(2, -1) = (0, 3) and T(4, 1) = (2, c)? Is T then unique?
  6. Find the linear map T: \mathbb{C} \to \mathbb{C}, with complex scalars, such that T(1 + i) = 2, and find T(i).

Proofs

  1. Let S, T \in \mathcal{L}(V, W) and let (\mathbf{v}_1, \dots, \mathbf{v}_m) span V. Prove that if S\mathbf{v}_k = T\mathbf{v}_k for every k, then S = T.
  2. Let T \in \mathcal{L}(V, W) and let U be a subspace of V. Prove that T(U) = \{T\mathbf{u} : \mathbf{u} \in U\} is a subspace of W.
  3. Let S, T \in \mathcal{L}(V) with ST = TS. Prove that (S + T)^2 = S^2 + 2ST + T^2. Then use Example 4 to express (D + M)^2 on \mathcal{P}(\mathbb{R}).
  4. With R the delay and B the advance on \mathbb{F}^{\infty}, prove that BR = I and RB \neq I.
  5. Let T \in \mathcal{L}(V) with T^2 = 0. Prove that (I + T)(I - T) = (I - T)(I + T) = I, and apply it to D on \mathcal{P}_1(\mathbb{R}).
  6. Let V and W be complex vector spaces and T: V \to W additive, with T(r\mathbf{v}) = rT\mathbf{v} for real r and T(i\mathbf{v}) = iT\mathbf{v} for all \mathbf{v}. Prove that T is linear.
  7. Let V and W be vector spaces over \mathbb{R} or \mathbb{C} and T: V \to W additive. Prove that T(q\mathbf{v}) = qT\mathbf{v} for every rational q.
Answers
  1. Evaluation at a point is linear, since (ap + q)(t) = ap(t) + q(t), so T(ap + q) = aTp + Tq and Result 1 applies. For p = 1 + 3x - x^2, p(2) = 1 + 6 - 4 = 3 and p(-1) = 1 - 3 - 1 = -3, so T(1 + 3x - x^2) = 6.
  2. (aA + C)^{T} = aA^{T} + C^{T} (The transpose and symmetric matrices), so T(aA + C) = aTA + TC. The value is \begin{bmatrix} 0 & 2 \\ -2 & 0 \end{bmatrix}.
  3. No: \det(2I_2) = 4 but 2\det I_2 = 2.
  4. No: for \mathbf{x} = (1, 1, 0, 0, \dots) the map gives (1, 0, 0, \dots), but for 2\mathbf{x} it gives (4, 0, 0, \dots), not (2, 0, 0, \dots).
  5. \int_0^1 x(af(x) + g(x))\,dx = a\int_0^1 xf(x)\,dx + \int_0^1 xg(x)\,dx by the linearity of the integral. Integrating by parts, T(e^x) = \big[xe^x - e^x\big]_0^1 = 0 - (-1) = 1.
  6. No: for \mathbf{x} = (1, 0, 0, \dots) the rectifier sends -\mathbf{x} to (1, 0, 0, \dots), but -1 times the image of \mathbf{x} is (-1, 0, 0, \dots).
  7. R^2\mathbf{x} = (0, 0, 3, 0, 6, 0, \dots), so the result is (3, 0, 5, 0, -2, 0, \dots).
  8. Mp = x + 2x^2 - x^3 and DMp = 1 + 4x - 3x^2; Dp = 2 - 2x and MDp = 2x - 2x^2; D^2Mp = 4 - 6x. And DMp - MDp = 1 + 2x - x^2 = p.
  9. ST(x, y) = S(x + y, 0) = (0, x + y); TS(x, y) = T(y, x) = (x + y, 0); (S + 2T)(x, y) = (2x + 3y, x). So ST \neq TS.
  10. (x, y) = (3x - y)(1, 2) + (y - 2x)(1, 3) with unique coefficients, so the two vectors form a basis. By Result 5, T(x, y) is (3x - y)(1, 0, 1) plus (y - 2x)(0, 1, 1), which is (3x - y, y - 2x, x), and T(5, 4) = (11, -6, 5).
  11. Only c = 7. The vectors (1, 1) and (2, -1) form a basis, so by Result 5 exactly one linear map has the first two values. Since (4, 1) = 2(1, 1) + (2, -1), Result 1 makes it send (4, 1) to 2(1, 2) + (0, 3) = (2, 7). For c = 7 that map is the only one; for any other c there is none.
  12. By Result 2, Tz = \lambda z with \lambda = T(1). Then \lambda(1 + i) = 2, so \lambda = \tfrac{2}{1 + i} = 1 - i, Tz = (1 - i)z, and T(i) = i - i^2 = 1 + i.
  13. Each \mathbf{v} \in V is a_1\mathbf{v}_1 + \dots + a_m\mathbf{v}_m for some scalars. By Result 1, S\mathbf{v} = \sum a_kS\mathbf{v}_k = \sum a_kT\mathbf{v}_k = T\mathbf{v}. So S and T agree at every vector, and S = T.
  14. Use the subspace test. \mathbf{0} = T\mathbf{0} \in T(U) by Result 1, since \mathbf{0} \in U. For \mathbf{u}, \mathbf{u}' \in U and a \in \mathbb{F}, T\mathbf{u} + T\mathbf{u}' = T(\mathbf{u} + \mathbf{u}') and aT\mathbf{u} = T(a\mathbf{u}), and \mathbf{u} + \mathbf{u}' and a\mathbf{u} lie in U, so both results lie in T(U).
  15. By the distributive laws, (S + T)^2 = (S + T)S + (S + T)T, which expands to S^2 + TS + ST + T^2. With TS = ST, the middle terms are 2ST (Axioms 5 and 8 in \mathcal{L}(V)). For D and M the expansion is D^2 + DM + MD + M^2 with MD = DM - I, so (D + M)^2 = D^2 + 2DM + M^2 - I: the binomial formula is off by I.
  16. BR(x_1, x_2, \dots) = B(0, x_1, x_2, \dots), which is (x_1, x_2, \dots), so BR = I. But RB(1, 0, 0, \dots) = R(0, 0, \dots) = \mathbf{0}, not (1, 0, 0, \dots), so RB \neq I.
  17. By Result 4, (I + T)(I - T) = I - T + T - T^2, which is I - T^2 = I; likewise (I - T)(I + T) = I + T - T - T^2 = I. On \mathcal{P}_1(\mathbb{R}), D^2 = 0, so I - D undoes I + D: (I + D)(a + bx) = (a + b) + bx, which I - D sends back to a + bx.
  18. Let c = r + si with r, s real. By Axioms 8 and 6, c\mathbf{v} = r\mathbf{v} + s(i\mathbf{v}). By additivity, T(c\mathbf{v}) = T(r\mathbf{v}) + T(s(i\mathbf{v})), and the two hypotheses turn this into rT\mathbf{v} + sT(i\mathbf{v}) and then into rT\mathbf{v} + siT\mathbf{v}, which is cT\mathbf{v}. With additivity, T is linear.
  19. The proof of T\mathbf{0} = \mathbf{0} in Result 1 used only additivity. For a positive integer n, T(n\mathbf{v}) = nT\mathbf{v} by induction, since (n + 1)\mathbf{v} = n\mathbf{v} + \mathbf{v} (Axioms 8 and 5). Also T\mathbf{v} + T(-\mathbf{v}) = T\mathbf{0} = \mathbf{0}, so T(-\mathbf{v}) = -T\mathbf{v}; with (-n)\mathbf{v} = -(n\mathbf{v}) and T\mathbf{0} = \mathbf{0}, the rule holds for every integer n. For q = m/n with m an integer and n \ge 1, nT(q\mathbf{v}) = T(nq\mathbf{v}) = T(m\mathbf{v}) = mT\mathbf{v}; scaling by \tfrac1n gives T(q\mathbf{v}) = \tfrac{m}{n}T\mathbf{v}.
math.wiki · CC-BY-SA-4.0