8.2 Linear Transformations and Their Matrices
Linear maps are the morphisms of linear algebra: they respect sums and scalar multiples, they are determined by what they do to a basis, and once bases are chosen each of them is a matrix. Rank–nullity, change of basis and similarity follow — together with duals, projections and nilpotent maps.
In §6.1 the derivative of $f:\mathbb{R}^n\to\mathbb{R}^m$ at a point was a linear map $Df(a)$, and the chain rule said that the derivative of a composition is the composition of the derivatives — computed as a product of Jacobian matrices. That is no accident. A matrix is nothing but a linear map written in coordinates, and matrix multiplication is defined so that it represents composition. Linear maps appear everywhere: differentiation and integration of polynomials, the trace, the transpose, the shift of a sequence, a rotation of the plane, the projection of a signal onto a few frequencies, the evaluation $p\mapsto p(2)$.
This section develops the theory of linear maps $T:V\to W$: kernel and image, the rank–nullity theorem and its consequences (injective $\Leftrightarrow$ surjective in equal finite dimension), isomorphisms, the algebra $\mathcal{L}(V,W)$, the matrix $[T]_{\mathcal{B}}^{\mathcal{C}}$ of a map, the change-of-basis formula $P^{-1}AP$ and similarity, the dual space, projections and nilpotent maps. CSIR NET asks about all of these: "the rank of the map $X\mapsto AX-XA$", "which of these maps are injective", "the matrix of $p\mapsto p'$ in this basis", "if $T^2=0$ then…". Notation from §8.1 is used throughout: $\mathbb{F}$ is a field, $\dim$ is dimension, $[v]_{\mathcal{B}}$ is a coordinate vector.
Linear Maps, Kernel and Image
Proof. Existence. Every $v\in V$ is $\sum c_ib_i$ for unique scalars $c_i$ (Theorem 8.6), so the formula defines a map. If $u=\sum c_ib_i$ and $v=\sum d_ib_i$, then $au+v=\sum(ac_i+d_i)b_i$, hence $T(au+v)=\sum(ac_i+d_i)w_i=aTu+Tv$. Clearly $Tb_i=w_i$. Uniqueness. If $S$ is linear with $Sb_i=w_i$, then $S\left(\sum c_ib_i\right)=\sum c_iSb_i=\sum c_iw_i$, so $S=T$. ∎
The values on a basis are completely free; the values on a dependent list are not. For instance, no linear $T:\mathbb{R}^2\to\mathbb{R}^2$ has $T(1,2)=(1,0)$ and $T(2,4)=(3,1)$: linearity forces $T(2,4)=2T(1,2)=(2,0)$.
Proof. (a) $T0=0$; if $Tu=Tv=0$ then $T(au+v)=0$. And $aTu+Tv=T(au+v)\in\operatorname{im}T$. (b) If $T$ is injective and $Tv=0=T0$, then $v=0$. Conversely, if $\ker T=\{0\}$ and $Tu=Tv$, then $T(u-v)=0$, so $u=v$. (c) Each $v$ is a combination $\sum a_is_i$, so $Tv=\sum a_iTs_i$. (d) If $\sum a_iTv_i=0$ then $T(\sum a_iv_i)=0$, so $\sum a_iv_i=0$ by (b), and independence of the $v_i$ gives $a_i=0$. For bijective $T$ combine this with (c). ∎
Proof. Let $u_1,\dots,u_k$ be a basis of $\ker T$ and extend it to a basis $u_1,\dots,u_k,v_1,\dots,v_r$ of $V$ (Theorem 8.9). We show that $Tv_1,\dots,Tv_r$ is a basis of $\operatorname{im}T$; then $\operatorname{rank}T=r=\dim V-k$. Spanning: by Theorem 8.17(c) the image is spanned by $Tu_1,\dots,Tu_k,Tv_1,\dots,Tv_r$, and $Tu_i=0$. Independence: if $\sum b_jTv_j=0$, then $T\left(\sum b_jv_j\right)=0$, so $\sum b_jv_j\in\ker T$ and $\sum b_jv_j=\sum a_iu_i$ for some scalars $a_i$. This is a linear relation among the basis vectors $u_1,\dots,u_k,v_1,\dots,v_r$ of $V$, so all $b_j=0$. ∎
Proof. (a) $\operatorname{rank}T\le\min(n,m)$, since $\operatorname{im}T\le W$ and $\operatorname{rank}T=n-\operatorname{nullity}T\le n$. If $n\gt m$, $\operatorname{nullity}T=n-\operatorname{rank}T\ge n-m\gt0$. If $n\lt m$, $\operatorname{rank}T\le n\lt m$, so $\operatorname{im}T\ne W$ (Theorem 8.9(4)). (b) $T$ is injective iff $\operatorname{nullity}T=0$, iff $\operatorname{rank}T=n=m$, iff $\operatorname{im}T=W$, using Theorem 8.17(b), rank–nullity and Theorem 8.9(4). (c) On $\mathbb{F}^{\mathbb{N}}$, the right shift $R(x_1,x_2,\dots)=(0,x_1,x_2,\dots)$ is injective but not surjective, and the left shift $L(x_1,x_2,\dots)=(x_2,x_3,\dots)$ is surjective but not injective. On $\mathbb{R}[x]$, $D:p\mapsto p'$ is surjective with kernel the constants, and $J:p\mapsto\int_0^xp(t)\,dt$ is injective with image $\{p:p(0)=0\}$. ∎
Isomorphisms and the Algebra $\mathcal{L}(V,W)$
Proof. (a) Let $w,w'\in W$ and put $v=T^{-1}w$, $v'=T^{-1}w'$. Then $T(av+v')=aw+w'$, so $T^{-1}(aw+w')=av+v'=aT^{-1}w+T^{-1}w'$. (b) It is linear by Theorem 8.6, injective since $[v]_{\mathcal{B}}=0$ means $v=0$, and surjective since every $(c_1,\dots,c_n)$ is the coordinate vector of $\sum c_ib_i$. (c) An isomorphism maps a basis to a basis (Theorem 8.17(d)), so equal dimension is necessary. Conversely, if both dimensions are $n$, compose the coordinate isomorphism $V\to\mathbb{F}^n$ with the inverse of the one for $W$. ∎
So, as vector spaces, $P_3(\mathbb{R})\cong\mathbb{R}^4\cong M_2(\mathbb{R})$, and $\operatorname{Sym}_3(\mathbb{R})\cong P_5(\mathbb{R})$ (both of dimension $6$). An isomorphism forgets every structure except addition and scaling — multiplication of polynomials or of matrices is not preserved.
Proof. (a) The vector space axioms hold pointwise, as for $\mathbb{F}^X$; linearity of $S+T$ and $aT$ is a one-line check. $ST(au+v)=S(aTu+Tv)=aSTu+STv$. Associativity holds for all maps of sets, and the bilinearity identities are checked by applying both sides to a vector, e.g. $S(T_1+T_2)v=S(T_1v+T_2v)=ST_1v+ST_2v$. (b) Fix bases $(b_j)$ of $V$ and $(c_i)$ of $W$ and let $E_{ij}$ be the linear map with $E_{ij}b_j=c_i$ and $E_{ij}b_l=0$ for $l\ne j$ (Theorem 8.16). If $Tb_j=\sum_ia_{ij}c_i$ then $T=\sum_{i,j}a_{ij}E_{ij}$, because both sides agree on the basis; and $\sum a_{ij}E_{ij}=0$ applied to $b_j$ gives $\sum_ia_{ij}c_i=0$, so all $a_{ij}=0$. The $mn$ maps $E_{ij}$ form a basis. (c) With a basis $b_1,b_2,\dots$, let $Sb_1=b_1$, $Sb_l=0$ ($l\ge2$) and $Tb_1=0$, $Tb_2=b_1$, $Tb_l=0$ ($l\ge3$). Then $STb_2=Sb_1=b_1$ but $TSb_2=0$, so $ST\ne TS$; and $TS=0$ since $TSb_1=Tb_1=0$ and $TSb_l=0$ for $l\ge2$. (d) If $Tv=0$ then $v=STv=0$, so $T$ is injective, hence bijective by Theorem 8.19(b). Then $S=S(TT^{-1})=(ST)T^{-1}=T^{-1}$, so $TS=TT^{-1}=I$. ∎
Part (d) is false in infinite dimension: the shifts of Theorem 8.19(c) satisfy $LR=I$ but $RL\ne I$ (it kills $x_1$), and $DJ=I$ on $\mathbb{R}[x]$ (fundamental theorem of calculus) while $JD\,p=p-p(0)$.
Proof. Well defined: if $u+\ker T=v+\ker T$, then $u-v\in\ker T$, so $Tu=Tv$. Linear: $a(u+\ker T)+(v+\ker T)=(au+v)+\ker T$ is sent to $T(au+v)=aTu+Tv$. Injective: $\bar T(v+\ker T)=0$ means $v\in\ker T$, i.e. $v+\ker T$ is the zero coset. Surjective: every $Tv\in\operatorname{im}T$ is $\bar T(v+\ker T)$. Finally $\dim V-\dim\ker T=\dim V/\ker T=\dim\operatorname{im}T$. ∎
The theorem is often the quickest way to identify a quotient: $C[0,1]/\{f:f(0)=0\}\cong\mathbb{R}$ because $f\mapsto f(0)$ is linear and onto $\mathbb{R}$ with exactly that kernel; and $M_n/\operatorname{Sym}_n\cong\operatorname{Skew}_n$ via $A\mapsto A-A^{T}$ (in $\mathbb{R}$).
The Matrix of a Linear Map
- $[Tv]_{\mathcal{C}}=[T]_{\mathcal{B}}^{\mathcal{C}}\,[v]_{\mathcal{B}}$ for every $v\in V$;
- $T\mapsto[T]_{\mathcal{B}}^{\mathcal{C}}$ is an isomorphism $\mathcal{L}(V,W)\to M_{m\times n}(\mathbb{F})$;
- if $S\in\mathcal{L}(W,X)$ and $\mathcal{D}$ is an ordered basis of $X$, then $[ST]_{\mathcal{B}}^{\mathcal{D}}=[S]_{\mathcal{C}}^{\mathcal{D}}\,[T]_{\mathcal{B}}^{\mathcal{C}}$;
- $T$ is an isomorphism iff $[T]_{\mathcal{B}}^{\mathcal{C}}$ is invertible (so $m=n$), and then $[T^{-1}]_{\mathcal{C}}^{\mathcal{B}}=\big([T]_{\mathcal{B}}^{\mathcal{C}}\big)^{-1}$;
- $\operatorname{rank}T$ equals the dimension of the column space of $[T]_{\mathcal{B}}^{\mathcal{C}}$.
Proof. Write $A=[T]_{\mathcal{B}}^{\mathcal{C}}$. (1) If $v=\sum_jx_jb_j$, then $$Tv=\sum_jx_j\sum_ia_{ij}c_i=\sum_i\Big(\sum_ja_{ij}x_j\Big)c_i,$$ and $\sum_ja_{ij}x_j$ is the $i$-th entry of $A[v]_{\mathcal{B}}$. (2) The $j$-th column of $[S+T]$ is $[(S+T)b_j]_{\mathcal{C}}=[Sb_j]_{\mathcal{C}}+[Tb_j]_{\mathcal{C}}$, and similarly for $aT$: the map is linear. It is bijective because, by Theorem 8.16, every choice of the $n$ columns (i.e. of the vectors $Tb_j$) comes from exactly one $T$. (3) The $j$-th column of $[ST]_{\mathcal{B}}^{\mathcal{D}}$ is $[S(Tb_j)]_{\mathcal{D}}=[S]_{\mathcal{C}}^{\mathcal{D}}[Tb_j]_{\mathcal{C}}$ by (1), which is $[S]_{\mathcal{C}}^{\mathcal{D}}$ times the $j$-th column of $A$ — the $j$-th column of the product. (4) If $T$ is invertible, (3) gives $[T^{-1}][T]=[I_V]_{\mathcal{B}}=I_n$ and $[T][T^{-1}]=I_m$, so $m=n$ and the matrices are inverse. Conversely, if $A$ has an inverse $B$, the map $S$ with $[S]_{\mathcal{C}}^{\mathcal{B}}=B$ (from (2)) satisfies $[ST]=I$ and $[TS]=I$, so $ST=I_V$, $TS=I_W$. (5) The coordinate isomorphism $w\mapsto[w]_{\mathcal{C}}$ carries $\operatorname{im}T=\operatorname{span}\{Tb_j\}$ onto the span of the columns of $A$, and isomorphisms preserve dimension. ∎
Two consequences are worth stating. First, matrix multiplication is associative because composition of maps is — no index juggling is needed. Second, for $A\in M_{m\times n}(\mathbb{F})$ the matrix of $L_A:x\mapsto Ax$ in the standard bases is $A$ itself, so everything said about maps applies to matrices: $\ker A$, $\operatorname{im}A$ (the column space), $\operatorname{rank}A$. That the column space and the row space have the same dimension is proved in §8.3.
- Apply $T$ to the first basis vector $b_1$ of the domain.
- Express $Tb_1$ in the basis $\mathcal{C}$ of the codomain; the coefficients form column $1$. (For $\mathcal{C}$ standard this is immediate; otherwise solve a small linear system.)
- Repeat for $b_2,\dots,b_n$. Columns, not rows: a common error is to write the coordinates of $Tb_j$ as the $j$-th row, which produces the transpose.
- Check with one vector: compute $Tv$ directly and as $[T][v]_{\mathcal{B}}$.
Change of Basis and Similarity
Proof. (b) $T=I_W\circ T\circ I_V$, so Theorem 8.23(3) gives $$[T]_{\mathcal{B}'}^{\mathcal{C}'}=[I_W]_{\mathcal{C}}^{\mathcal{C}'}\,[T]_{\mathcal{B}}^{\mathcal{C}}\,[I_V]_{\mathcal{B}'}^{\mathcal{B}},$$ which is $Q^{-1}[T]_{\mathcal{B}}^{\mathcal{C}}P$. (a) is the case $W=V$, $\mathcal{C}=\mathcal{B}$, $\mathcal{C}'=\mathcal{B}'$, $Q=P$. In words: to apply $T$ in new coordinates, convert to old coordinates ($P$), apply $T$ there, and convert back ($P^{-1}$). ∎
Proof. (a) $A=I^{-1}AI$; if $B=P^{-1}AP$ then $A=(P^{-1})^{-1}BP^{-1}$; if also $C=Q^{-1}BQ$ then $C=(PQ)^{-1}A(PQ)$. (b) If $A=[T]_{\mathcal{B}}$ and $B=[T]_{\mathcal{B}'}$, Theorem 8.24 gives $B=P^{-1}AP$. Conversely, if $B=P^{-1}AP$, the columns $p_1,\dots,p_n$ of $P$ are independent (Theorem 8.23(4)), hence a basis $\mathcal{B}'$ of $\mathbb{F}^n$ with transition matrix $P$ to the standard basis $\mathcal{E}$; so $[L_A]_{\mathcal{B}'}=P^{-1}[L_A]_{\mathcal{E}}P=B$. (c) The rank is the rank of the operator (Theorem 8.23(5)), which does not involve a basis. For the trace, $\operatorname{tr}(XY)=\sum_{i,j}x_{ij}y_{ji}=\operatorname{tr}(YX)$, so $$\operatorname{tr}(P^{-1}AP)=\operatorname{tr}(APP^{-1})=\operatorname{tr}A.$$ The determinant and characteristic polynomial follow from the product rule for determinants (§8.3): $$\det(xI-P^{-1}AP)=\det\big(P^{-1}(xI-A)P\big),$$ which equals $\det(xI-A)$. Next, $(P^{-1}AP)^k=P^{-1}A^kP$ (the inner factors $PP^{-1}$ cancel), and adding multiples gives $p(P^{-1}AP)=P^{-1}p(A)P$. Finally $B^{T}=P^{T}A^{T}(P^{-1})^{T}=Q^{-1}A^{T}Q$ with $Q=(P^{T})^{-1}$. (d) $P^{-1}IP=I$, so the only matrix similar to $I$ is $I$. ∎
Proof. As in the proof of rank–nullity, choose a basis $v_1,\dots,v_r,u_1,\dots,u_k$ of $V$ with $u_1,\dots,u_k$ a basis of $\ker T$; then $Tv_1,\dots,Tv_r$ is a basis of $\operatorname{im}T$, which we extend to a basis $\mathcal{C}$ of $W$. With $\mathcal{B}=(v_1,\dots,v_r,u_1,\dots,u_k)$, $Tv_j$ is the $j$-th vector of $\mathcal{C}$ and $Tu_i=0$, which is the stated matrix. If $A$ has rank $r$, applying this to $L_A$ and Theorem 8.24(b) shows $A$ is equivalent to the normal form; equivalence is an equivalence relation, so matrices of equal rank are equivalent. Conversely, $QAP$ is the matrix of $L_A$ in other bases, so it has the same rank as $A$. ∎
So equivalence (two independent changes of basis, on the domain and the codomain) is a coarse relation with one invariant, the rank; similarity (the same change of basis on both sides) is much finer, and classifying matrices up to similarity is the subject of Chapter 9 (diagonal, triangular and Jordan forms).
Dual Spaces, Projections and Nilpotent Maps
The dual space of $V$ is $V^*=\mathcal{L}(V,\mathbb{F})$, the space of linear functionals on $V$. If $\mathcal{B}=(b_1,\dots,b_n)$ is a basis of $V$, the dual basis $\mathcal{B}^*=(f_1,\dots,f_n)$ is defined by $f_i(b_j)=\delta_{ij}$ ($1$ if $i=j$, $0$ otherwise); $f_i$ reads off the $i$-th coordinate. For $W\le V$, the annihilator is $W^0=\{f\in V^*:f(w)=0\text{ for all }w\in W\}$.
For $T\in\mathcal{L}(V,W)$, the dual (transpose) map $T^{t}:W^*\to V^*$ is $T^{t}(g)=g\circ T$.
Proof. (a) Each $f_i$ exists and is unique by Theorem 8.16. The functionals $f$ and $\sum_if(b_i)f_i$ agree on every $b_j$, hence are equal: the $f_i$ span. If $\sum a_if_i=0$, evaluating at $b_j$ gives $a_j=0$: they are independent. (b) Extend a basis $b_1,\dots,b_k$ of $W$ to a basis $b_1,\dots,b_n$ of $V$, with dual basis $f_1,\dots,f_n$. A functional $f=\sum f(b_i)f_i$ vanishes on $W$ iff $f(b_1)=\cdots=f(b_k)=0$, i.e. iff $f\in\operatorname{span}\{f_{k+1},\dots,f_n\}$; so $\dim W^0=n-k$. (c) $\hat v$ is linear in $f$, and $v\mapsto\hat v$ is linear in $v$. If $v\ne0$, extend $v$ to a basis $v=b_1,b_2,\dots,b_n$; the first dual functional has $f_1(v)=1$, so $\hat v\ne0$. The map is injective between spaces of the same dimension $n=\dim V^*=\dim V^{**}$, hence an isomorphism. (d) Let $\mathcal{C}=(c_i)$ with dual basis $(g_i)$, and $A=[T]_{\mathcal{B}}^{\mathcal{C}}$. Then $$T^{t}(g_i)(b_j)=g_i(Tb_j)=g_i\Big(\sum_ka_{kj}c_k\Big)=a_{ij},$$ so by (a), $T^{t}(g_i)=\sum_ja_{ij}f_j$: the $i$-th column of $[T^{t}]$ is the $i$-th row of $A$. ∎
In infinite dimension $V^*$ is strictly "larger": for $V=\mathbb{R}[x]$ (countable dimension) the dual $V^*\cong\mathbb{R}^{\mathbb{N}}$ (a functional is any choice of values on $1,x,x^2,\dots$) has uncountable dimension, and $V\to V^{**}$ is injective but not onto. Part (d) is the conceptual reason why $\operatorname{rank}A=\operatorname{rank}A^{T}$ (§8.3).
Proof. (a) Decompositions are unique, so $E$ is well defined; if $v=u+w$ and $v'=u'+w'$ then $av+v'=(au+u')+(aw+w')$, so $E(av+v')=aEv+Ev'$. $Eu=u$ for $u\in U$ gives $E^2v=E(u)=u=Ev$ and $\operatorname{im}E=U$; $Ev=0$ iff $u=0$ iff $v\in W$. (b) Every $v$ equals $Ev+(v-Ev)$ with $Ev\in\operatorname{im}E$ and $E(v-Ev)=Ev-E^2v=0$. If $v\in\operatorname{im}E\cap\ker E$, write $v=Ew$; then $v=Ew=E^2w=Ev=0$. So $V=\operatorname{im}E\oplus\ker E$; on $\operatorname{im}E$, $E(Ew)=Ew$, and $E$ kills $\ker E$ — that is the projection onto $\operatorname{im}E$ along $\ker E$. For $I-E$: $(I-E)^2=I-2E+E^2=I-E$; $(I-E)v=0$ iff $v=Ev$ iff $v\in\operatorname{im}E$; and $(I-E)v=v-Ev$ runs exactly over $\ker E$ (if $Ev=0$ then $v=(I-E)v$). (c) Combine a basis of $\operatorname{im}E$ ($r$ vectors, fixed by $E$) with a basis of $\ker E$ (killed by $E$). The trace is a similarity invariant (Theorem 8.25), so $\operatorname{tr}E=r$ in any basis. ∎
- $\{0\}\subsetneq\ker N\subsetneq\ker N^2\subsetneq\cdots\subsetneq\ker N^k=V$; hence $k\le n$ and $N^n=0$;
- the jumps $d_j=\dim\ker N^j-\dim\ker N^{j-1}$ satisfy $d_1\ge d_2\ge\cdots\ge d_k\ge1$; in particular, if $\operatorname{rank}N=n-1$ then $k=n$;
- $N$ is not invertible, and $I-N$ is invertible with $(I-N)^{-1}=I+N+\cdots+N^{k-1}$.
Proof. (1) $N^jv=0$ implies $N^{j+1}v=0$, so the kernels increase. If $\ker N^j=\ker N^{j+1}$ for some $j$, then also $\ker N^{j+1}=\ker N^{j+2}$: if $N^{j+2}v=0$ then $Nv\in\ker N^{j+1}=\ker N^j$, so $N^{j+1}v=0$. Thus once two consecutive kernels agree, all later ones agree. Since $\ker N^k=V\ne\ker N^{k-1}$ (by minimality of $k$), the chain must increase strictly at every step up to $k$. Each strict step raises the dimension by at least $1$, from $\dim\ker N^0=0$ to $n$, so $k\le n$; and $\ker N\ne\{0\}$. (2) Define $\varphi:\ker N^{j+1}/\ker N^j\to\ker N^j/\ker N^{j-1}$ by $\varphi(v+\ker N^j)=Nv+\ker N^{j-1}$. If $v\in\ker N^{j+1}$ then $Nv\in\ker N^j$; if $v\in\ker N^j$ then $Nv\in\ker N^{j-1}$ — so $\varphi$ is well defined, and it is linear. It is injective: $Nv\in\ker N^{j-1}$ means $N^jv=0$, i.e. $v\in\ker N^j$. Hence $d_{j+1}\le d_j$ (Theorem 8.14). If $\operatorname{rank}N=n-1$ then $d_1=1$, so every $d_j=1$ and reaching dimension $n$ takes $n$ steps. (3) $N$ has a nonzero kernel. And $(I-N)(I+N+\cdots+N^{k-1})=I-N^k=I$; the two factors commute, so this is a two-sided inverse. ∎
- $T0=0$ is necessary, not sufficient: translations fail it; $(x,y)\mapsto\sqrt[3]{x^3+y^3}$ passes it and is even homogeneous, yet is not additive.
- Linearity depends on the field: $z\mapsto\bar z$ is $\mathbb{R}$-linear, not $\mathbb{C}$-linear.
- Injective $\Leftrightarrow$ surjective needs equal finite dimension: the shifts on $\mathbb{F}^{\mathbb{N}}$, and $D$, $J$ on $\mathbb{R}[x]$, separate the two properties.
- A one-sided inverse is two-sided only in finite dimension: $LR=I\ne RL$ for the shifts; $DJ=I\ne JD$.
- $\ker T$ and $\operatorname{im}T$ need not be complementary: for $N=E_{12}$, $\ker N=\operatorname{im}N$. $V=\ker T\oplus\operatorname{im}T$ holds for projections, not in general.
- $\mathcal{L}(V)$ is not commutative and has zero divisors: $E_{11}E_{21}=0$ while $E_{21}E_{11}=E_{21}\ne0$.
- Same rank, trace, determinant and characteristic polynomial do not imply similar: $I_2$ versus $\begin{pmatrix}1&1\\0&1\end{pmatrix}$.
- Similar matrices need not commute or look alike: $\operatorname{diag}(1,2)\sim\begin{pmatrix}1&1\\0&2\end{pmatrix}$, but the two do not commute.
- Columns, not rows: the $j$-th column of $[T]_{\mathcal{B}}^{\mathcal{C}}$ is $[Tb_j]_{\mathcal{C}}$; writing rows gives the transpose (the matrix of the dual map). The matrix also depends on the order of the basis.
- The direction of the transition matrix: with $P$ having the new basis vectors as columns, $[v]_{\text{old}}=P[v]_{\text{new}}$ and $[T]_{\text{new}}=P^{-1}[T]_{\text{old}}P$ — not $PAP^{-1}$.
- Sums and products: a sum of nilpotents need not be nilpotent; a product of projections need not be a projection; $2E$ is never a projection unless $E=0$ (over $\mathbb{R}$).
- The determinant is not linear (Example 1(c)); the trace is, and $\operatorname{tr}(AB)=\operatorname{tr}(BA)$, but $\operatorname{tr}(ABC)\ne\operatorname{tr}(BAC)$ in general.
- Rank–nullity counts the domain: $\operatorname{rank}+\operatorname{nullity}=\dim V$, never $\dim W$; and it needs $\dim V\lt\infty$.
Exercises
- 1 Which are linear? (a) $T(x,y)=(x+y,x-y)$; (b) $T(x,y)=(xy,x)$; (c) $T(p)=xp(x)$ on $\mathbb{R}[x]$; (d) $T(A)=A^{T}$ on $M_n$; (e) $T(f)=f(0)^2$ on $C[0,1]$; (f) $T(f)=\int_0^1f(t)\,dt$ on $C[0,1]$.
- 2 Find bases of the kernel and the image of $T:\mathbb{R}^4\to\mathbb{R}^3$, $T(x)=(x_1+x_2,\ x_2+x_3,\ x_3+x_4)$.
- 3 Find the kernel, the rank and the nullity of $T:P_3(\mathbb{R})\to\mathbb{R}^2$, $T(p)=(p(0),p(1))$.
- 4 For $T:M_n(\mathbb{R})\to M_n(\mathbb{R})$, $T(A)=A-A^{T}$, identify $\ker T$ and $\operatorname{im}T$ and verify rank–nullity.
- 5 Show that no linear operator on $\mathbb{R}^3$ satisfies $\ker T=\operatorname{im}T$, and find one on $\mathbb{R}^2$ that does.
- 6 Show that $D:p\mapsto p'$ on $P_n(\mathbb{R})$ has rank $n$, and that $T(p)=p''+p'$ on $P_4(\mathbb{R})$ has rank $4$. What are their images?
- 7 Let $T:V\to W$ with $V$ finite-dimensional. Prove that $T$ is injective iff $\operatorname{rank}T=\dim V$, and that $\operatorname{rank}T\le\min(\dim V,\dim W)$.
- 8 Let $T\in\mathcal{L}(V)$, $\dim V\lt\infty$. Prove that $\dim\ker T^2\le2\dim\ker T$.
- 1 Write down explicit isomorphisms $P_3(\mathbb{R})\to\mathbb{R}^4$ and $\mathbb{R}^4\to M_2(\mathbb{R})$.
- 2 Which pairs are isomorphic over $\mathbb{R}$? (a) $\operatorname{Sym}_3(\mathbb{R})$ and $P_5(\mathbb{R})$; (b) $\mathbb{C}^2$ and $\mathbb{R}^4$; (c) $\{A\in M_3(\mathbb{R}):\operatorname{tr}A=0\}$ and $\mathbb{R}^8$; (d) $\mathbb{R}^3$ and $P_3(\mathbb{R})$.
- 3 Find $\dim\mathcal{L}(\mathbb{R}^3,P_2(\mathbb{R}))$ and $\dim\mathcal{L}(M_2(\mathbb{R}),\mathbb{R})$.
- 4 Find $A,B\in M_2(\mathbb{R})$ with $AB=0$ but $BA\ne0$.
- 5 Prove that if $\dim V\lt\infty$ and $S,T\in\mathcal{L}(V)$ with $ST$ invertible, then both $S$ and $T$ are invertible. Show by an example on $\mathbb{R}^{\mathbb{N}}$ that this fails in infinite dimension.
- 6 Use the first isomorphism theorem to show $C[0,1]/W\cong\mathbb{R}^2$ for $W=\{f:f(0)=f(1)=0\}$.
- 7 Let $T:V\to W$ be linear, $V$ finite-dimensional, and $Y\le W$. Prove that $T^{-1}(Y)=\{v:Tv\in Y\}$ is a subspace with $\dim T^{-1}(Y)=\dim(Y\cap\operatorname{im}T)+\operatorname{nullity}T$.
- 1 Find the matrix of $T(p)(x)=p(x+1)$ on $P_2(\mathbb{R})$ relative to $(1,x,x^2)$.
- 2 Let $T:\mathbb{R}^3\to\mathbb{R}^2$, $T(x,y,z)=(x+y,\ y+z)$. Find $[T]$ in the standard bases and $[T]_{\mathcal{B}}^{\mathcal{C}}$ for $\mathcal{B}=((1,0,0),(1,1,0),(1,1,1))$ and $\mathcal{C}=((1,0),(1,1))$.
- 3 Let $R_\theta$ be the rotation of $\mathbb{R}^2$ through $\theta$. Use $R_\theta R_\varphi=R_{\theta+\varphi}$ and Theorem 8.23 to derive the addition formulas for $\cos(\theta+\varphi)$ and $\sin(\theta+\varphi)$.
- 4 Find the standard matrix of the reflection of $\mathbb{R}^2$ in the line through $0$ and $(1,3)$.
- 5 Find the standard matrix of the orthogonal projection of $\mathbb{R}^3$ onto the plane $x+y+z=0$ (that is, the projection onto that plane along the line spanned by $(1,1,1)$).
- 6 Find an invertible $P$ with $P^{-1}\begin{pmatrix}1&2\\0&3\end{pmatrix}P=\begin{pmatrix}1&0\\0&3\end{pmatrix}$.
- 7 Show that $\begin{pmatrix}a&b\\0&a\end{pmatrix}$ with $b\ne0$ is similar to $\begin{pmatrix}a&1\\0&a\end{pmatrix}$ but not to $aI$.
- 8 In $P_2(\mathbb{R})$ let $\mathcal{B}=(1,x,x^2)$ and $\mathcal{B}'=(1,x-1,(x-1)^2)$. Find the transition matrix $P$ from $\mathcal{B}'$ to $\mathcal{B}$, and use it to find the coordinates of $3+2x+x^2$ in $\mathcal{B}'$.
- 1 Find the dual basis of $(1,0,1),(0,1,1),(1,1,0)$ in $(\mathbb{R}^3)^*$.
- 2 Let $W\le V$, $\dim V=n$. Prove that $W^{00}$ (the annihilator of $W^0$ in $V^{**}$) corresponds to $W$ under $v\mapsto\hat v$.
- 3 Prove that $E$ is a projection iff $I-E$ is, and that then $E(I-E)=0$.
- 4 Show that a real $3\times3$ matrix $E$ with $E^2=E$ and $\operatorname{tr}E=0$ is the zero matrix, and that the only invertible projection is $I$.
- 5 For $N=\begin{pmatrix}0&1&0\\0&0&1\\0&0&0\end{pmatrix}$ find the ranks of $N,N^2,N^3$ and the index of nilpotency, and compute $(I-N)^{-1}$.
- 6 Prove that the only nilpotent projection is $0$, and that a nilpotent operator on an $n$-dimensional space satisfies $N^n=0$ — so a $4\times4$ matrix with $N^{10}=0$ has $N^4=0$.
- 7 Give a nilpotent $4\times4$ matrix of index $2$ and rank $2$, and explain why no nilpotent $4\times4$ matrix has index $2$ and rank $3$.
- 8 Let $T\in\mathcal{L}(V)$ with $T^2=T$ and $S\in\mathcal{L}(V)$. Prove: $ST=TS$ iff $S$ maps $\operatorname{im}T$ into itself and $\ker T$ into itself.
- 9 Applied. A graphics engine first rotates every point of the plane by $30^\circ$ anticlockwise and then projects it orthogonally onto the $x$-axis (the screen line). Find the matrix of the composite map, its rank, and its kernel, and describe which points end up at the origin.
E1. CSIR Part B GATE MA Let $V$ be the space of real polynomials of degree at most $4$ and $T:V\to V$, $T(p)=p''+p'$. Then (A) $\operatorname{rank}T=3$; (B) $\operatorname{rank}T=4$; (C) $\operatorname{nullity}T=2$; (D) $T$ is invertible.
E2. CSIR Part B The linear operator $T$ on $\mathbb{R}^2$ has standard matrix $\begin{pmatrix}1&2\\3&4\end{pmatrix}$. Its matrix relative to the ordered basis $b_1=(1,1)$, $b_2=(1,-1)$ is (A) $\begin{pmatrix}5&-1\\-2&0\end{pmatrix}$; (B) $\begin{pmatrix}5&-2\\-1&0\end{pmatrix}$; (C) $\begin{pmatrix}3&-1\\7&-1\end{pmatrix}$; (D) $\begin{pmatrix}4&2\\3&1\end{pmatrix}$.
E3. CSIR Part B The number of surjective linear maps from $\mathbb{F}_2^3$ to $\mathbb{F}_2^2$ is (A) $42$; (B) $48$; (C) $56$; (D) $64$.
E4. CSIR Part C GATE MA Let $V$ be a finite-dimensional real vector space, $\dim V=n$, and $T\in\mathcal{L}(V)$. Which of the following are true? (One or more options may be correct.) (A) If $T^2=0$ then $\operatorname{rank}T\le n/2$. (B) If $T$ is injective then $T$ is surjective. (C) $\ker T\cap\operatorname{im}T=\{0\}$. (D) If $T^2=T$ then $V=\ker T\oplus\operatorname{im}T$.
E5. CSIR Part C On $V=\mathbb{R}[x]$ let $D(p)=p'$ and $J(p)(x)=\int_0^xp(t)\,dt$. Which of the following are true? (A) $DJ=I$. (B) $JD=I$. (C) $D$ is surjective but not injective. (D) $J$ is injective but not surjective.
E6. CSIR Part C Let $A,B\in M_n(\mathbb{R})$ be similar. Which of the following are necessarily true? (A) $\operatorname{tr}A=\operatorname{tr}B$. (B) $A^{T}$ and $B^{T}$ are similar. (C) $AB=BA$. (D) Every matrix $C\in M_n(\mathbb{R})$ with the same trace, determinant and rank as $A$ is similar to $A$.
Suggested plan — 2 lectures × 60 min (plus 30 min for duals and nilpotents if time allows).
- Maps, kernels, rank–nullity, isomorphisms. Definition and Example 1 as a quick poll (include the cube-root map). Theorem 8.16 — stress "values on a basis are free". Kernel and image, rank–nullity with full proof, Theorem 8.19 with the infinite-dimensional counterexamples on the board. Example 2 (commutator map) worked in full. Project Interactive 1: start with the rank-1 matrix and play the coset sweep, then the nilpotent one ($\ker=\operatorname{im}$), then the custom matrix as $d\to4$. Close with Theorem 8.22 and the dimension count of $\mathcal{L}(V,W)$.
- Matrices and change of basis. Definition 8.13 — "columns are images" — with Example 4. Theorem 8.23 (products = compositions). Transition matrices and $P^{-1}AP$, Example 5 done both ways. Project Interactive 2 for the symmetric operator, press Eigen-basis, then try the shear and the rotation to see that no basis diagonalises them — a teaser for Chapter 9. Theorem 8.25 and the invariants; E2 and E6 as in-class problems.
Discussion questions. (1) Why must the matrix of $T$ have the coordinates of $Tb_j$ as columns for $[ST]=[S][T]$ to hold? (2) Rank alone decides equivalence; why can it not decide similarity? Which extra invariants have we met? (3) In which results of this section is finite dimension essential, and what is the counterexample each time?
Misconceptions to watch for. Writing rows instead of columns; using $PAP^{-1}$ with $P$ built from the new basis; assuming $V=\ker T\oplus\operatorname{im}T$; thinking $T(0)=0$ or homogeneity proves linearity; believing that matrices with equal trace, determinant and rank are similar; forgetting that rank–nullity is about the domain.
Suggested homework. I(2, 3, 4, 6, 8), II(2, 5, 6), III(2, 4, 6, 8), IV(1, 4, 5, 7, 9); Exam corner E1, E3, E5.
Research question. A digital signal is a vector of samples, written in the standard basis. Is there a basis in which most of its coordinates are almost zero — so that it can be stored with far fewer numbers?
Why it matters. Image, audio and video compression are changes of basis followed by throwing away small coordinates. JPEG uses a cosine basis; JPEG 2000 — used in digital cinema and in medical imaging archives — and the fingerprint-compression standard adopted by police agencies use wavelet bases. The simplest wavelet basis, found by Alfréd Haar in 1909, already shows the whole idea.
What to do.
- In $\mathbb{R}^8$ take the Haar basis: $h_0=\tfrac1{\sqrt8}(1,\dots,1)$; $h_1=\tfrac1{\sqrt8}(1,1,1,1,-1,-1,-1,-1)$; $h_2,h_3$ equal to $\tfrac12(1,1,-1,-1)$ on one half and $0$ on the other; and $h_4,\dots,h_7$ equal to $\tfrac1{\sqrt2}(1,-1)$ on one pair of places and $0$ elsewhere. Check that these vectors are orthonormal, hence a basis (Chapter 10), and write the transition matrix $P$ whose columns they are. Since $P^{-1}=P^{T}$, the Haar coordinates of a signal $s$ are $c=P^{T}s$.
- Compute $c$ for the signal $s=(10,12,11,9,4,2,3,5)$.
- Keep only the two largest coordinates (in absolute value), set the others to $0$, and reconstruct $\tilde s=Pc_{\text{kept}}$. Compute the error $\|s-\tilde s\|$ and compare it with keeping the two largest samples of $s$ itself.
- Explain, using Theorem 8.24, why "compress in one basis, display in another" costs only two matrix multiplications, and why an orthonormal transition matrix makes the error easy to predict.
Tools: Python with NumPy (or a spreadsheet); for images, the PyWavelets package.
Hand in: Your table of Haar coordinates, the two reconstructions with their errors, and a one-paragraph explanation of why piecewise-smooth signals have few large Haar coordinates.
Going further. Apply the two-dimensional Haar transform to a small grey-scale image and plot the error against the number of coefficients kept; read about the discrete cosine transform used by JPEG.
🧪 Self-check quiz
Q1. $T:\mathbb{R}^7\to\mathbb{R}^4$ is linear and surjective. Then $\dim\ker T$ is
Q2. Which map $\mathbb{R}^2\to\mathbb{R}^2$ is linear?
Q3. The matrix of $D=d/dx$ on $P_2(\mathbb{R})$ in the basis $(1,x,x^2)$ is
Q4. A real $4\times4$ matrix $E$ satisfies $E^2=E$ and has rank $2$. Then $\operatorname{tr}E$ is
Q5. $\dim\mathcal{L}(\mathbb{R}^3,\mathbb{R}^2)$ equals