Chapter 8 · Vector Spaces and Linear Transformations

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

📘 Definition 8.9 — Linear transformation
Let $V,W$ be vector spaces over $\mathbb{F}$. A map $T:V\to W$ is linear (a linear transformation, or linear operator when $W=V$) if $$T(u+v)=Tu+Tv\quad\text{and}\quad T(av)=a\,Tv$$ for all $u,v\in V$, $a\in\mathbb{F}$; equivalently $T(au+v)=aTu+Tv$. Then $T0=0$ and $T\left(\sum a_iv_i\right)=\sum a_iTv_i$. A linear map $V\to\mathbb{F}$ is a linear functional.
✏️ Example 1 — Linear or not?
Decide which maps are linear: (a) $L_A:\mathbb{F}^n\to\mathbb{F}^m$, $x\mapsto Ax$ for a fixed $A\in M_{m\times n}(\mathbb{F})$; (b) $D:\mathbb{R}[x]\to\mathbb{R}[x]$, $p\mapsto p'$; (c) $\operatorname{tr}:M_n\to\mathbb{F}$ and $\det:M_n\to\mathbb{F}$ ($n\ge2$); (d) $T(x,y)=(x+1,y)$ on $\mathbb{R}^2$; (e) $z\mapsto\bar z$ on $\mathbb{C}$, viewed over $\mathbb{R}$ and over $\mathbb{C}$; (f) $f(x,y)=\sqrt[3]{x^3+y^3}$ on $\mathbb{R}^2$.
(a), (b) Linear $A(ax+y)=aAx+Ay$ is the distributive law for matrices; $(ap+q)'=ap'+q'$ is the linearity of the derivative.
(c) The trace is linear: $\operatorname{tr}(aA+B)=a\operatorname{tr}A+\operatorname{tr}B$. The determinant is not: $\det(2I)=2^n\ne2\det I$, and $\det(E_{11}+(I-E_{11}))=1$ while $\det E_{11}+\det(I-E_{11})=0$.
(d) Not linear $T(0,0)=(1,0)\ne(0,0)$. A translation moves the origin, and every linear map fixes it.
(e) Depends on the field $\overline{z+w}=\bar z+\bar w$ always and $\overline{az}=a\bar z$ for real $a$: it is $\mathbb{R}$-linear. But $\overline{i\cdot1}=-i\ne i\cdot\bar1$, so it is not $\mathbb{C}$-linear.
(f) Not linear It is homogeneous ($f(tx,ty)=tf(x,y)$ for all real $t$) but not additive: $f(1,0)+f(0,1)=2$ while $f(1,1)=\sqrt[3]2$. Homogeneity alone does not give linearity.
📐 Theorem 8.16 — A linear map is determined by its values on a basis
Let $b_1,\dots,b_n$ be a basis of $V$ and let $w_1,\dots,w_n\in W$ be arbitrary. Then there is exactly one linear map $T:V\to W$ with $Tb_i=w_i$ for $i=1,\dots,n$, namely $T\left(\sum c_ib_i\right)=\sum c_iw_i$. (The same holds for an infinite basis.)

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 $T(2,4)=2T(1,2)=(2,0)$.

📘 Definition 8.10 — Kernel, image, nullity, rank
For linear $T:V\to W$, the kernel (null space) is $\ker T=\{v\in V:Tv=0\}$ and the image (range) is $\operatorname{im}T=\{Tv:v\in V\}$. When finite, $\operatorname{nullity}T=\dim\ker T$ and $\operatorname{rank}T=\dim\operatorname{im}T$.
📐 Theorem 8.17 — Basic properties of kernel and image
Let $T:V\to W$ be linear. (a) $\ker T\le V$ and $\operatorname{im}T\le W$. (b) $T$ is injective if and only if $\ker T=\{0\}$. (c) If $S$ spans $V$, then $T(S)$ spans $\operatorname{im}T$; so $\operatorname{im}T=\operatorname{span}\{Tb_1,\dots,Tb_n\}$ for a basis $(b_i)$. (d) If $T$ is injective, it maps independent sets to independent sets; if $T$ is bijective, it maps bases to bases.

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). ∎

📐 Theorem 8.18 — The rank–nullity theorem
Let $V$ be finite-dimensional and $T:V\to W$ linear ($W$ arbitrary). Then $\operatorname{im}T$ is finite-dimensional and $$\dim V=\operatorname{rank}T+\operatorname{nullity}T.$$

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 $b_j=0$. ∎

📐 Theorem 8.19 — Injective, surjective, bijective
Let $T:V\to W$ be linear with $\dim V=n$ and $\dim W=m$ finite. (a) If $n\gt m$, $T$ is not injective; if $n\lt m$, $T$ is not surjective. (b) If $n=m$, then $T$ is injective $\iff$ $T$ is surjective $\iff$ $T$ is bijective. (c) Both statements fail when $V=W$ is infinite-dimensional.

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 $\{p:p(0)=0\}$. ∎

✏️ Example 2 — The commutator map $X\mapsto AX-XA$
Let $A=\begin{pmatrix}1&1\\0&1\end{pmatrix}$ and $T:M_2(\mathbb{R})\to M_2(\mathbb{R})$, $T(X)=AX-XA$. Show that $T$ is linear, find $\ker T$ and $\operatorname{im}T$ with bases, and check rank–nullity.
Linear $T(aX+Y)=A(aX+Y)-(aX+Y)A=a(AX-XA)+(AY-YA)$.
Compute For $X=\begin{pmatrix}p&q\\r&s\end{pmatrix}$: $AX=\begin{pmatrix}p+r&q+s\\r&s\end{pmatrix}$, $XA=\begin{pmatrix}p&p+q\\r&r+s\end{pmatrix}$, so $$T(X)=\begin{pmatrix}r&s-p\\0&-r\end{pmatrix}.$$
Kernel and image $T(X)=0\iff r=0,\ s=p$: $\ker T=\{pI+qE_{12}\}$, basis $I,E_{12}$, nullity $2$ — the matrices commuting with $A$. The image is $\left\{\begin{pmatrix}r&t\\0&-r\end{pmatrix}\right\}$ ($t=s-p$ is arbitrary), basis $E_{11}-E_{22},\ E_{12}$, rank $2$.
Check $2+2=4=\dim M_2$. Note that $E_{12}$ lies in both the kernel and the image: $\ker T\cap\operatorname{im}T\ne\{0\}$, so $M_2\ne\ker T\oplus\operatorname{im}T$ here. Every image matrix has trace $0$, as it must: $\operatorname{tr}(AX)=\operatorname{tr}(XA)$.
✏️ Example 3 — Which linear maps exist? Prove or disprove
Does there exist a linear map (a) $T:\mathbb{R}^5\to\mathbb{R}^2$ with $\dim\ker T=2$? (b) $T:\mathbb{R}^3\to\mathbb{R}^3$ with $\ker T=\operatorname{im}T$? (c) $T:\mathbb{R}^4\to\mathbb{R}^4$ with $\ker T=\operatorname{im}T$? (d) $T:\mathbb{R}^2\to\mathbb{R}^3$ with $T(1,1)=(1,0,2)$ and $T(1,-1)=(3,2,0)$ — and if so, what is $T(x,y)$?
(a) No $\operatorname{rank}T\le2$, so $\operatorname{nullity}T=5-\operatorname{rank}T\ge3$.
(b) No $\ker T=\operatorname{im}T$ would give rank $=$ nullity $=k$ with $2k=3$.
(c) Yes By Theorem 8.16 define $Te_1=Te_2=0$, $Te_3=e_1$, $Te_4=e_2$. Then $\operatorname{im}T=\operatorname{span}\{e_1,e_2\}$, and $T(x_1,x_2,x_3,x_4)=(x_3,x_4,0,0)$ vanishes iff $x_3=x_4=0$, so $\ker T=\operatorname{span}\{e_1,e_2\}$ as well. Such a $T$ satisfies $T^2=0$.
(d) Yes, uniquely $(1,1),(1,-1)$ is a basis of $\mathbb{R}^2$, so Theorem 8.16 applies. $e_1=\tfrac12[(1,1)+(1,-1)]$ and $e_2=\tfrac12[(1,1)-(1,-1)]$ give $Te_1=(2,1,1)$, $Te_2=(-1,-1,1)$, so $$T(x,y)=(2x-y,\ x-y,\ x+y).$$ Check: $T(1,1)=(1,0,2)$ and $T(1,-1)=(3,2,0)$. ✓
Interactive 1 — Rank–Nullity in the Plane: Cosets of the Kernel Collapse to Points

Left: a grid of vectors $x$ in the domain. Right: their images $Ax$. For a singular matrix the red kernel line is crushed to $0$, and every line parallel to it — a coset $x+\ker A$ — is crushed to a single point of the blue image line; dots of one colour share one image. Press Play to slide the amber coset across the plane and watch its image move along $\operatorname{im}A$: the cosets of the kernel correspond one-to-one to the points of the image ($V/\ker T\cong\operatorname{im}T$, Theorem 8.22). For an invertible matrix the amber object is a point on a circle and its image runs round an ellipse.

2.50
A—
det A—
ker A—
im A—
rank + nullity—

Isomorphisms and the Algebra $\mathcal{L}(V,W)$

📘 Definition 8.11 — Isomorphism
A bijective linear map $T:V\to W$ is an isomorphism; $V$ and $W$ are isomorphic, $V\cong W$, if an isomorphism exists. An isomorphism $V\to V$ is an automorphism; the automorphisms of $V$ form the general linear group $GL(V)$.
📐 Theorem 8.20 — Finite-dimensional spaces are classified by their dimension
(a) The inverse of an isomorphism is linear. (b) If $\mathcal{B}$ is an ordered basis of $V$ with $n$ vectors, the coordinate map $v\mapsto[v]_{\mathcal{B}}$ is an isomorphism $V\to\mathbb{F}^n$. (c) Finite-dimensional spaces $V,W$ over $\mathbb{F}$ are isomorphic if and only if $\dim V=\dim W$.

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.

📘 Definition 8.12 — The space $\mathcal{L}(V,W)$ and the algebra $\mathcal{L}(V)$
$\mathcal{L}(V,W)$ is the set of all linear maps $V\to W$, with $(S+T)v=Sv+Tv$ and $(aT)v=a(Tv)$; $\mathcal{L}(V)=\mathcal{L}(V,V)$. For $T\in\mathcal{L}(V,W)$ and $S\in\mathcal{L}(W,X)$ the composition $ST=S\circ T\in\mathcal{L}(V,X)$; powers $T^k$ and polynomials $p(T)=\sum a_kT^k$ ($T^0=I$) are defined for $T\in\mathcal{L}(V)$.
📐 Theorem 8.21 — Properties of $\mathcal{L}(V,W)$
(a) $\mathcal{L}(V,W)$ is a vector space; compositions of linear maps are linear; composition is associative and bilinear: $S(T_1+T_2)=ST_1+ST_2$, $(S_1+S_2)T=S_1T+S_2T$, $a(ST)=(aS)T=S(aT)$. Thus $\mathcal{L}(V)$ is an associative algebra with identity $I$. (b) If $\dim V=n$ and $\dim W=m$, then $\dim\mathcal{L}(V,W)=mn$. (c) If $\dim V\ge2$, $\mathcal{L}(V)$ is not commutative and has zero divisors ($ST=0$ with $S,T\ne0$). (d) If $\dim V\lt\infty$ and $S,T\in\mathcal{L}(V)$ satisfy $ST=I$, then $T$ and $S$ are invertible and $TS=I$.

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 $JD\,p=p-p(0)$.

📐 Theorem 8.22 — The first isomorphism theorem
Let $T:V\to W$ be linear. Then $\bar T:V/\ker T\to\operatorname{im}T$, $\ \bar T(v+\ker T)=Tv$, is a well-defined isomorphism: $$V/\ker T\cong\operatorname{im}T.$$ When $\dim V\lt\infty$, taking dimensions (Theorem 8.14) gives rank–nullity again.

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

📘 Definition 8.13 — The matrix of $T$ relative to two bases
Let $\mathcal{B}=(b_1,\dots,b_n)$ be an ordered basis of $V$, $\mathcal{C}=(c_1,\dots,c_m)$ one of $W$, and $T\in\mathcal{L}(V,W)$. Write $Tb_j=\sum_{i=1}^ma_{ij}c_i$. The $m\times n$ matrix $[T]_{\mathcal{B}}^{\mathcal{C}}=(a_{ij})$ is the matrix of $T$ relative to $\mathcal{B}$ and $\mathcal{C}$: its $j$-th column is the coordinate vector $[Tb_j]_{\mathcal{C}}$. For $T\in\mathcal{L}(V)$ we write $[T]_{\mathcal{B}}=[T]_{\mathcal{B}}^{\mathcal{B}}$.
📐 Theorem 8.23 — Matrices represent maps; products represent compositions
With notation as in Definition 8.13:
  1. $[Tv]_{\mathcal{C}}=[T]_{\mathcal{B}}^{\mathcal{C}}\,[v]_{\mathcal{B}}$ for every $v\in V$;
  2. $T\mapsto[T]_{\mathcal{B}}^{\mathcal{C}}$ is an isomorphism $\mathcal{L}(V,W)\to M_{m\times n}(\mathbb{F})$;
  3. if $S\in\mathcal{L}(W,X)$ and $\mathcal{D}$ is an ordered basis of $X$, then $[ST]_{\mathcal{B}}^{\mathcal{D}}=[S]_{\mathcal{C}}^{\mathcal{D}}\,[T]_{\mathcal{B}}^{\mathcal{C}}$;
  4. $T$ is an isomorphism iff $[T]_{\mathcal{B}}^{\mathcal{C}}$ is invertible (so $m=n$), and then $[T^{-1}]_{\mathcal{C}}^{\mathcal{B}}=\big([T]_{\mathcal{B}}^{\mathcal{C}}\big)^{-1}$;
  5. $\operatorname{rank}T$ 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.

🧭 Problem-solving strategy — writing down $[T]_{\mathcal{B}}^{\mathcal{C}}$
  1. Apply $T$ to the first basis vector $b_1$ of the domain.
  2. 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.)
  3. 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.
  4. Check with one vector: compute $Tv$ directly and as $[T][v]_{\mathcal{B}}$.
✏️ Example 4 — Matrices of differentiation and of a difference operator
On $P_3(\mathbb{R})$ with the basis $\mathcal{B}=(1,x,x^2,x^3)$: (a) find $[D]_{\mathcal{B}}$ for $D=d/dx$ and deduce $D^4=0$; (b) find $[D]_{\mathcal{B}'}$ for $\mathcal{B}'=\left(1,x,\tfrac{x^2}2,\tfrac{x^3}6\right)$; (c) find the matrix, the rank and the kernel of $\Delta:p(x)\mapsto p(x+1)-p(x)$.
(a) $D1=0$, $Dx=1$, $Dx^2=2x$, $Dx^3=3x^2$. The columns are the coordinate vectors: $$[D]_{\mathcal{B}}=\begin{pmatrix}0&1&0&0\\0&0&2&0\\0&0&0&3\\0&0&0&0\end{pmatrix}.$$ Each application of $D$ lowers the degree, so $D^4p=0$ for all $p\in P_3$, and $[D]_{\mathcal{B}}^4=[D^4]_{\mathcal{B}}=0$ by Theorem 8.23(3).
(b) $D\left(\tfrac{x^2}2\right)=x$ and $D\left(\tfrac{x^3}6\right)=\tfrac{x^2}2$: each basis vector goes to the previous one. So $[D]_{\mathcal{B}'}$ has $1,1,1$ on the superdiagonal and zeros elsewhere — the nilpotent Jordan block of size $4$ (Chapter 9).
(c) $\Delta1=0$, $\Delta x=1$, $\Delta x^2=2x+1$, $\Delta x^3=3x^2+3x+1$: $$[\Delta]_{\mathcal{B}}=\begin{pmatrix}0&1&1&1\\0&0&2&3\\0&0&0&3\\0&0&0&0\end{pmatrix}.$$ The last three columns are independent (triangular with nonzero "diagonal" $1,2,3$), so $\operatorname{rank}\Delta=3$ and $\operatorname{im}\Delta=P_2$; $\ker\Delta$ is the constants (a polynomial with $p(x+1)=p(x)$ for all $x$ takes the value $p(0)$ at every integer, so $p-p(0)$ has infinitely many roots). Rank–nullity: $3+1=4$. ✓

Change of Basis and Similarity

📘 Definition 8.14 — Transition matrix
Let $\mathcal{B}=(b_1,\dots,b_n)$ and $\mathcal{B}'=(b'_1,\dots,b'_n)$ be ordered bases of $V$. The transition matrix from $\mathcal{B}'$ to $\mathcal{B}$ is $P=[I]_{\mathcal{B}'}^{\mathcal{B}}$: its $j$-th column is $[b'_j]_{\mathcal{B}}$, the new basis vector written in the old basis. By Theorem 8.23, $P$ is invertible, $P^{-1}=[I]_{\mathcal{B}}^{\mathcal{B}'}$, and $$[v]_{\mathcal{B}}=P\,[v]_{\mathcal{B}'}.$$
📐 Theorem 8.24 — The change-of-basis formula
(a) If $T\in\mathcal{L}(V)$ and $P$ is the transition matrix from $\mathcal{B}'$ to $\mathcal{B}$, then $$[T]_{\mathcal{B}'}=P^{-1}\,[T]_{\mathcal{B}}\,P.$$ (b) More generally, for $T\in\mathcal{L}(V,W)$, bases $\mathcal{B},\mathcal{B}'$ of $V$ with transition matrix $P$ and $\mathcal{C},\mathcal{C}'$ of $W$ with transition matrix $Q$ (from $\mathcal{C}'$ to $\mathcal{C}$): $[T]_{\mathcal{B}'}^{\mathcal{C}'}=Q^{-1}[T]_{\mathcal{B}}^{\mathcal{C}}P$.

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 ($P^{-1}$). ∎

📘 Definition 8.15 — Similar and equivalent matrices
$A,B\in M_n(\mathbb{F})$ are similar, $A\sim B$, if $B=P^{-1}AP$ for some invertible $P\in M_n(\mathbb{F})$. $A,B\in M_{m\times n}(\mathbb{F})$ are equivalent if $B=QAP$ for some invertible $Q\in M_m(\mathbb{F})$ and $P\in M_n(\mathbb{F})$.
📐 Theorem 8.25 — Similarity: what it means and what it preserves
(a) Similarity is an equivalence relation on $M_n(\mathbb{F})$. (b) $A\sim B$ if and only if $A$ and $B$ are the matrices of one linear operator on an $n$-dimensional space relative to two ordered bases. (c) Similar matrices have the same rank, trace, determinant and characteristic polynomial, and $A^k\sim B^k$, $p(A)\sim p(B)$ for every polynomial $p$, and $A^{T}\sim B^{T}$. (d) The converse of (c) is false: $I_2$ and $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ share rank, trace, determinant and characteristic polynomial but are not similar.

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 $I$. ∎

📐 Theorem 8.26 — Rank normal form; equivalence is decided by rank
If $T\in\mathcal{L}(V,W)$ has rank $r$ ($V,W$ finite-dimensional), there are bases $\mathcal{B}$ of $V$ and $\mathcal{C}$ of $W$ with $$[T]_{\mathcal{B}}^{\mathcal{C}}=\begin{pmatrix}I_r&0\\0&0\end{pmatrix}.$$ Consequently two $m\times n$ matrices are equivalent if and only if they have the same rank.

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).

✏️ Example 5 — A reflection, computed through a better basis
Find the standard matrix $A$ of the reflection $T$ of $\mathbb{R}^2$ in the line $y=2x$.
Adapted basis Let $b'_1=(1,2)$ (on the line) and $b'_2=(-2,1)$ (perpendicular to it). Then $Tb'_1=b'_1$ and $Tb'_2=-b'_2$, so $[T]_{\mathcal{B}'}=\operatorname{diag}(1,-1)$.
Transition matrix $P=\begin{pmatrix}1&-2\\2&1\end{pmatrix}$ (columns $b'_1,b'_2$ in standard coordinates), $P^{-1}=\tfrac15\begin{pmatrix}1&2\\-2&1\end{pmatrix}$.
Change basis By Theorem 8.24, $A=P\,\operatorname{diag}(1,-1)\,P^{-1}$, which gives $$A=\frac15\begin{pmatrix}-3&4\\4&3\end{pmatrix}.$$
Check $A(1,2)^{T}=\tfrac15(5,10)^{T}=(1,2)^{T}$ and $A(-2,1)^{T}=\tfrac15(10,-5)^{T}=(2,-1)^{T}$. Also $\operatorname{tr}A=0$ and $\det A=-1$, as for $\operatorname{diag}(1,-1)$ (Theorem 8.25(c)), and $A^2=I$.
Interactive 2 — One Operator, Many Matrices: $[T]_{\mathcal{B}'}=P^{-1}AP$

Choose an operator $T$ of the plane (standard matrix $A$) and a basis $\mathcal{B}'=(b_1,b_2)$ by the two angles. The dashed arrows are $Tb_1$ and $Tb_2$; the matrix on the right is $[T]_{\mathcal{B}'}$ — its first column is $Tb_1$ written in the basis $b_1,b_2$ (the thin parallelogram shows the decomposition). Press Play to rotate $b_1$: the entries change all the time, but the trace and determinant never do (Theorem 8.25). Press Eigen-basis to jump to a basis in which the matrix is diagonal — when one exists.

b₁, b₂ angles—
[T]_B′ column 1—
[T]_B′ column 2—
trace · det—
diagonal?—

Dual Spaces, Projections and Nilpotent Maps

📘 Definition 8.16 — Dual space, dual basis, annihilator, transpose

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 $T^{t}(g)=g\circ T$.

📐 Theorem 8.27 — Duality in finite dimension
Let $\dim V=n$. (a) The dual basis is a basis of $V^*$, and $f=\sum_if(b_i)f_i$ for every $f\in V^*$; so $\dim V^*=n$. (b) $\dim W^0=n-\dim W$ for every $W\le V$. (c) The map $V\to V^{**}$, $v\mapsto\hat v$ with $\hat v(f)=f(v)$, is an isomorphism that needs no choice of basis. (d) $[T^{t}]_{\mathcal{C}^*}^{\mathcal{B}^*}=\big([T]_{\mathcal{B}}^{\mathcal{C}}\big)^{T}$, the transpose matrix.

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).

✏️ Example 6 — A dual basis and an annihilator
(a) Find the dual basis of $b_1=(1,1)$, $b_2=(1,2)$ in $(\mathbb{R}^2)^*$. (b) In $P_2(\mathbb{R})$, show that the evaluations $\varepsilon_0,\varepsilon_1,\varepsilon_2$ ($\varepsilon_a(p)=p(a)$) form a basis of $P_2^*$ and find the basis of $P_2$ to which they are dual. (c) Find a basis of $W^0$ for $W=\operatorname{span}\{(1,1,0,0),(0,1,1,0)\}\le\mathbb{R}^4$.
(a) Rows of the inverse If $B=\begin{pmatrix}1&1\\1&2\end{pmatrix}$ has columns $b_1,b_2$, the conditions $f_i(b_j)=\delta_{ij}$ say that the rows of the matrix of $(f_1,f_2)$ times $B$ give $I$: these rows form $B^{-1}=\begin{pmatrix}2&-1\\-1&1\end{pmatrix}$. So $f_1(x,y)=2x-y$, $f_2(x,y)=-x+y$. Check: $f_1(1,1)=1$, $f_1(1,2)=0$, $f_2(1,1)=0$, $f_2(1,2)=1$. ✓
(b) Lagrange polynomials We need $L_0,L_1,L_2\in P_2$ with $L_i(j)=\delta_{ij}$: $$L_0=\tfrac12(x-1)(x-2),\quad L_1=-x(x-2),\quad L_2=\tfrac12x(x-1).$$ They are independent (apply $\varepsilon_j$ to a relation), so they are a basis of $P_2$, and by construction $\varepsilon_0,\varepsilon_1,\varepsilon_2$ is its dual basis. Every $p\in P_2$ is $p(0)L_0+p(1)L_1+p(2)L_2$ — the Lagrange interpolation formula is Theorem 8.27(a) in disguise.
(c) $f=a_1x_1+a_2x_2+a_3x_3+a_4x_4$ annihilates $W$ iff $a_1+a_2=0$ and $a_2+a_3=0$. So $W^0$ has basis $x_1-x_2+x_3$ and $x_4$; $\dim W^0=2=4-2$, as Theorem 8.27(b) predicts.
📘 Definition 8.17 — Projection
$E\in\mathcal{L}(V)$ is a projection (idempotent) if $E^2=E$. If $V=U\oplus W$, the projection onto $U$ along $W$ is $E(u+w)=u$ ($u\in U$, $w\in W$).
📐 Theorem 8.28 — Projections and direct sums
(a) If $V=U\oplus W$, the projection onto $U$ along $W$ is linear, $E^2=E$, $\operatorname{im}E=U$ and $\ker E=W$. (b) Conversely, if $E^2=E$, then $V=\operatorname{im}E\oplus\ker E$, $E$ is the identity on $\operatorname{im}E$, and $E$ is the projection onto $\operatorname{im}E$ along $\ker E$. Also $I-E$ is a projection, with $\operatorname{im}(I-E)=\ker E$ and $\ker(I-E)=\operatorname{im}E$. (c) If $\dim V=n$ and $E^2=E$ has rank $r$, then in a suitable basis $[E]=\operatorname{diag}(I_r,0)$; hence $\operatorname{tr}E=r\cdot1$, and $\operatorname{rank}E=\operatorname{tr}E$ over $\mathbb{R}$ or $\mathbb{C}$.

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. ∎

📘 Definition 8.18 — Nilpotent operator
$N\in\mathcal{L}(V)$ (or $N\in M_n(\mathbb{F})$) is nilpotent if $N^k=0$ for some $k\ge1$; the least such $k$ is the index of nilpotency.
📐 Theorem 8.29 — Structure of nilpotent maps
Let $N\in\mathcal{L}(V)$ be nilpotent of index $k$, with $\dim V=n\ge1$. Then:
  1. $\{0\}\subsetneq\ker N\subsetneq\ker N^2\subsetneq\cdots\subsetneq\ker N^k=V$; hence $k\le n$ and $N^n=0$;
  2. 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 $k=n$;
  3. $N$ 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. ∎

✏️ Example 7 — A projection, and four claims about nilpotents and projections
(a) Show that $E=\begin{pmatrix}1&0&1\\0&1&1\\0&0&0\end{pmatrix}$ is a projection, find its image and kernel, and check $\operatorname{rank}E=\operatorname{tr}E$. (b) Prove or disprove: (i) a sum of two nilpotent matrices is nilpotent; (ii) a sum of two commuting nilpotent matrices is nilpotent; (iii) a product of two projections is a projection; (iv) a nilpotent $5\times5$ matrix of rank $4$ has index $5$.
(a) $E(x,y,z)=(x+z,\,y+z,\,0)$, and applying $E$ again to $(x+z,y+z,0)$ returns it unchanged: $E^2=E$. The image is the plane $z=0$ (it contains $e_1=Ee_1$ and $e_2=Ee_2$), the kernel is the line spanned by $(1,1,-1)$. So $E$ projects onto the plane $z=0$ along the (non-perpendicular) direction $(1,1,-1)$. $\operatorname{rank}E=2=1+1+0=\operatorname{tr}E$. ✓
(i) False $N_1=E_{12}$ and $N_2=E_{21}$ are nilpotent ($N_i^2=0$), but $(N_1+N_2)^2=\begin{pmatrix}0&1\\1&0\end{pmatrix}^2=I$.
(ii) True If $MN=NM$ with $M^p=0$ and $N^q=0$, the binomial theorem applies: $(M+N)^{p+q-1}=\sum_i\binom{p+q-1}{i}M^iN^{p+q-1-i}$, and in every term either $i\ge p$ or $p+q-1-i\ge q$.
(iii) False; (iv) True (iii) $E_1=\begin{pmatrix}1&0\\0&0\end{pmatrix}$ and $E_2=\tfrac12\begin{pmatrix}1&1\\1&1\end{pmatrix}$ are projections, but $E_1E_2=\tfrac12\begin{pmatrix}1&1\\0&0\end{pmatrix}$ has square $\tfrac14\begin{pmatrix}1&1\\0&0\end{pmatrix}\ne E_1E_2$. (If $E_1E_2=E_2E_1$, the product is a projection: $(E_1E_2)^2=E_1^2E_2^2$.) (iv) By Theorem 8.29(2), rank $4=n-1$ forces index $n=5$.
⚠️ Counterexamples and pitfalls
  1. $T0=0$ 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.
  2. Linearity depends on the field: $z\mapsto\bar z$ is $\mathbb{R}$-linear, not $\mathbb{C}$-linear.
  3. 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.
  4. A one-sided inverse is two-sided only in finite dimension: $LR=I\ne RL$ for the shifts; $DJ=I\ne JD$.
  5. $\ker T$ 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.
  6. $\mathcal{L}(V)$ is not commutative and has zero divisors: $E_{11}E_{21}=0$ while $E_{21}E_{11}=E_{21}\ne0$.
  7. Same rank, trace, determinant and characteristic polynomial do not imply similar: $I_2$ versus $\begin{pmatrix}1&1\\0&1\end{pmatrix}$.
  8. 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.
  9. 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.
  10. 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 $PAP^{-1}$.
  11. 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 $\mathbb{R}$).
  12. 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.
  13. Rank–nullity counts the domain: $\operatorname{rank}+\operatorname{nullity}=\dim V$, never $\dim W$; and it needs $\dim V\lt\infty$.

Exercises

I
Linear maps, kernel, image and rank–nullity.
  • 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 (a), (c), (d), (f) are linear. (b) is not: $T(2,2)=(4,2)\ne2T(1,1)=(2,2)$. (e) is not: $T(2\cdot1)=4\ne2T(1)$.
2 $T(x)=0$ gives $x_2=-x_1$, $x_3=x_1$, $x_4=-x_1$: $\ker T=\operatorname{span}\{(1,-1,1,-1)\}$, nullity $1$. Rank $=4-1=3$, so $\operatorname{im}T=\mathbb{R}^3$ (basis $e_1,e_2,e_3$; e.g. $T(e_1)=e_1$, $T(e_4)=e_3$, $T(e_2)-T(e_1)=e_2$).
3 $T(1)=(1,1)$ and $T(x)=(0,1)$ span $\mathbb{R}^2$: rank $2$, nullity $4-2=2$. $\ker T=\{p:p(0)=p(1)=0\}=\{x(x-1)(a+bx)\}$, basis $x^2-x$, $x^3-x^2$.
4 $\ker T=\operatorname{Sym}_n$ (dimension $\tfrac{n(n+1)}2$). $\operatorname{im}T\subseteq\operatorname{Skew}_n$, and every skew $S$ equals $T(\tfrac12S)$, so $\operatorname{im}T=\operatorname{Skew}_n$ (dimension $\tfrac{n(n-1)}2$). Sum $n^2$. ✓
5 rank $=$ nullity $=k$ would need $2k=3$. On $\mathbb{R}^2$: $T(x,y)=(y,0)$, whose kernel and image are both the $x$-axis.
6 $\ker D$ is the constants, so $\operatorname{rank}D=(n+1)-1=n$ and $\operatorname{im}D=P_{n-1}$ (it lies in $P_{n-1}$ and has its dimension). For $T=D^2+D$: if $p''+p'=0$ and $\deg p=d\ge1$, then $p'$ has degree $d-1$ while $p''$ has smaller degree — impossible; so $\ker T$ is the constants and $\operatorname{rank}T=5-1=4$, $\operatorname{im}T=P_3$.
7 $T$ injective $\iff\operatorname{nullity}T=0\iff\operatorname{rank}T=\dim V$. $\operatorname{rank}T\le\dim V$ by rank–nullity and $\le\dim W$ because $\operatorname{im}T\le W$.
8 Restrict $T$ to $K=\ker T^2$. It maps $K$ into $\ker T$ (if $T^2v=0$ then $T(Tv)=0$), and its kernel is $K\cap\ker T=\ker T$. Rank–nullity for $T|_K$: $\dim K=\dim\ker T+\dim T(K)\le\dim\ker T+\dim\ker T$.
II
Isomorphisms, $\mathcal{L}(V,W)$ and the first isomorphism theorem.
  • 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 $a_0+a_1x+a_2x^2+a_3x^3\mapsto(a_0,a_1,a_2,a_3)$ (coordinates in $1,x,x^2,x^3$), and $(a,b,c,d)\mapsto\begin{pmatrix}a&b\\c&d\end{pmatrix}$. Both are linear bijections.
2 Compare real dimensions (Theorem 8.20): (a) $6=6$, yes; (b) $4=4$, yes; (c) $9-1=8$, yes; (d) $3\ne4$, no.
3 $3\cdot3=9$ and $4\cdot1=4$ (Theorem 8.21(b)).
4 $A=E_{11}$, $B=E_{21}$: $AB=E_{11}E_{21}=0$, while $BA=E_{21}E_{11}=E_{21}\ne0$.
5 Let $U=(ST)^{-1}$. Then $S(TU)=I$, so by Theorem 8.21(d) $S$ is invertible; then $T=S^{-1}(ST)$ is a product of invertible maps. In $\mathbb{R}^{\mathbb{N}}$: $LR=I$ is invertible, but neither $L$ (not injective) nor $R$ (not surjective) is.
6 $\Phi(f)=(f(0),f(1))$ is linear $C[0,1]\to\mathbb{R}^2$, onto (take $f(x)=a+(b-a)x$), with kernel $W$. Theorem 8.22 gives $C[0,1]/W\cong\mathbb{R}^2$.
7 $T^{-1}(Y)$ contains $0$ and is closed under $au+v$ since $T(au+v)=aTu+Tv\in Y$. Apply rank–nullity to $T$ restricted to $T^{-1}(Y)$: its kernel is $\ker T$ (contained in $T^{-1}(Y)$) and its image is $Y\cap\operatorname{im}T$ (any $y=Tv\in Y$ has $v\in T^{-1}(Y)$).
III
Matrices of linear maps and change of basis.
  • 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 $T1=1$, $Tx=1+x$, $Tx^2=1+2x+x^2$: $[T]=\begin{pmatrix}1&1&1\\0&1&2\\0&0&1\end{pmatrix}$ (Pascal's triangle in the columns).
2 Standard: $\begin{pmatrix}1&1&0\\0&1&1\end{pmatrix}$. Now $T(1,0,0)=(1,0)=1\cdot(1,0)+0\cdot(1,1)$, $T(1,1,0)=(2,1)=(1,0)+(1,1)$, $T(1,1,1)=(2,2)=2(1,1)$. So $[T]_{\mathcal{B}}^{\mathcal{C}}=\begin{pmatrix}1&1&0\\0&1&2\end{pmatrix}$; this agrees with $Q^{-1}AP$ for $Q=\begin{pmatrix}1&1\\0&1\end{pmatrix}$, $P=\begin{pmatrix}1&1&1\\0&1&1\\0&0&1\end{pmatrix}$.
3 $[R_\theta]=\begin{pmatrix}\cos\theta&-\sin\theta\\\sin\theta&\cos\theta\end{pmatrix}$ (columns $R_\theta e_1$, $R_\theta e_2$). Multiplying $[R_\theta][R_\varphi]$ and comparing with $[R_{\theta+\varphi}]$ entry by entry gives $\cos(\theta+\varphi)=\cos\theta\cos\varphi-\sin\theta\sin\varphi$ and $\sin(\theta+\varphi)=\sin\theta\cos\varphi+\cos\theta\sin\varphi$.
4 As in Example 5 with $b'_1=(1,3)$, $b'_2=(-3,1)$: $A=\tfrac1{10}\begin{pmatrix}-8&6\\6&8\end{pmatrix}=\begin{pmatrix}-0.8&0.6\\0.6&0.8\end{pmatrix}$. Check: $A(1,3)^{T}=(1,3)^{T}$.
5 $E=I-F$ where $F$ projects onto the line: $Fv=\tfrac{v\cdot(1,1,1)}{3}(1,1,1)$, $[F]=\tfrac13J$. So $[E]=\tfrac13\begin{pmatrix}2&-1&-1\\-1&2&-1\\-1&-1&2\end{pmatrix}$; trace $2=$ rank, and $E(1,1,1)=0$.
6 We need a basis $p_1,p_2$ with $Ap_1=p_1$ and $Ap_2=3p_2$: $p_1=(1,0)$, and $(A-3I)p_2=0$ gives $p_2=(1,1)$. $P=\begin{pmatrix}1&1\\0&1\end{pmatrix}$, and indeed $P^{-1}AP=\operatorname{diag}(1,3)$.
7 With $P=\operatorname{diag}(1,1/b)$: $P^{-1}\begin{pmatrix}a&b\\0&a\end{pmatrix}P=\begin{pmatrix}a&1\\0&a\end{pmatrix}$ (in the basis $e_1,\ e_2/b$, $T(e_2/b)=e_1+a\,e_2/b$). Only $aI$ is similar to $aI$, and $b\ne0$.
8 Columns of $P$ are $1$, $x-1$, $(x-1)^2=1-2x+x^2$ in $\mathcal{B}$: $P=\begin{pmatrix}1&-1&1\\0&1&-2\\0&0&1\end{pmatrix}$. Then $[p]_{\mathcal{B}'}=P^{-1}(3,2,1)^{T}=(6,4,1)^{T}$: $3+2x+x^2=6+4(x-1)+(x-1)^2$ (the Taylor expansion at $1$: $p(1)=6$, $p'(1)=4$, $\tfrac12p''(1)=1$).
IV
Dual spaces, projections and nilpotent maps.
  • 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.
1 Put the vectors as the columns of $B$; the dual functionals are the rows of $B^{-1}=\tfrac12\begin{pmatrix}1&-1&1\\-1&1&1\\1&1&-1\end{pmatrix}$: $f_1=\tfrac12(x-y+z)$, $f_2=\tfrac12(-x+y+z)$, $f_3=\tfrac12(x+y-z)$. Check $f_1(1,0,1)=1$, $f_1(0,1,1)=0$, $f_1(1,1,0)=0$.
2 For $w\in W$ and $f\in W^0$, $\hat w(f)=f(w)=0$, so $\hat W\subseteq W^{00}$. Dimensions: $\dim W^{00}=n-\dim W^0=n-(n-\dim W)=\dim W=\dim\hat W$ (Theorem 8.27(b), applied in $V^*$). So $\hat W=W^{00}$.
3 $(I-E)^2=I-2E+E^2$ equals $I-E$ iff $E^2=E$. Then $E(I-E)=E-E^2=0$.
4 $\operatorname{rank}E=\operatorname{tr}E=0$ (Theorem 8.28(c)), so $E=0$. If $E^2=E$ and $E$ is invertible, multiply by $E^{-1}$: $E=I$.
5 $N^2=E_{13}$, $N^3=0$: ranks $2,1,0$, index $3$ (as Theorem 8.29(2) predicts, since rank $N=2=n-1$). $(I-N)^{-1}=I+N+N^2=\begin{pmatrix}1&1&1\\0&1&1\\0&0&1\end{pmatrix}$.
6 If $E^2=E$ and $E^k=0$, then $E=E^2=\cdots=E^k=0$. The second claim is Theorem 8.29(1); $N^{10}=0$ makes $N$ nilpotent, and then $N^4=0$.
7 $N=E_{13}+E_{24}$: $N^2=0$, rank $2$. Index $2$ means $N^2=0$, i.e. $\operatorname{im}N\subseteq\ker N$, so $\operatorname{rank}N\le\operatorname{nullity}N=4-\operatorname{rank}N$, i.e. $\operatorname{rank}N\le2$.
8 If $ST=TS$: for $v=Tw$, $Sv=STw=T(Sw)\in\operatorname{im}T$; for $v\in\ker T$, $T(Sv)=S(Tv)=0$. Conversely, write $v=Tv+(v-Tv)$ (Theorem 8.28(b)). Then $S(Tv)\in\operatorname{im}T$ is fixed by $T$, and $S(v-Tv)\in\ker T$ is killed by $T$; so $TSv=S(Tv)+0=STv$.
9 Rotation $R=\begin{pmatrix}\frac{\sqrt3}2&-\frac12\\\frac12&\frac{\sqrt3}2\end{pmatrix}$, projection $P_x=\begin{pmatrix}1&0\\0&0\end{pmatrix}$, composite (rotate first, so on the right) $$P_xR=\begin{pmatrix}\frac{\sqrt3}2&-\frac12\\0&0\end{pmatrix}.$$ Rank $1$. Kernel: $\tfrac{\sqrt3}2x=\tfrac12y$, i.e. the line $y=\sqrt3x$ at $60^\circ$ — exactly the points that the rotation carries onto the $y$-axis (to $90^\circ$), which the projection then sends to $0$.
🎯 Exam corner CSIR Part BCSIR Part CGATE MA

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.

Idea Find the kernel by a degree argument, then use rank–nullity.
Work If $\deg p=d\ge1$, then $p'+p''$ has degree exactly $d-1$ (the leading term of $p'$ cannot be cancelled by $p''$), so it is nonzero. Hence $\ker T$ is the constants: nullity $1$, rank $5-1=4$. (A) would need nullity $2$ — that is $p\mapsto p''$; (D) fails because constants are killed.
Answer (B).

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) $\begin{pmatrix}4&2\\3&1\end{pmatrix}$.

Idea Column $j$ is $Tb_j$ in the basis $b_1,b_2$.
Work $Tb_1=(3,7)=5b_1-2b_2$; $Tb_2=(-1,-1)=-b_1+0\cdot b_2$. So the columns are $(5,-2)$ and $(-1,0)$. Check: trace $5=1+4$ and determinant $-2=4-6$. (B) is the transpose (rows instead of columns); (C) is $AP$, the images in standard coordinates; (D) has trace $5$ but determinant $-2$ — still not it, because its first column would mean $Tb_1=4b_1+3b_2=(7,1)$.
Answer (A).

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) $64$.

Idea Linear maps $\leftrightarrow$ $2\times3$ matrices over $\mathbb{F}_2$ ($2^6=64$ of them); a map is onto iff its three columns span $\mathbb{F}_2^2$.
Work The columns fail to span iff all of them lie in one line $\{0,v\}$. For each of the $3$ lines there are $2^3=8$ such matrices, and the zero matrix is counted three times: $3\cdot8-2=22$ non-surjective maps. So $64-22=42$. (Equivalently: the two rows must be independent, $(8-1)(8-2)=42$.)
Answer (A).

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 $V=\ker T\oplus\operatorname{im}T$.

(A) True $T^2=0$ means $\operatorname{im}T\subseteq\ker T$, so $r\le n-r$.
(B) True Theorem 8.19(b).
(C) False $T=E_{12}$ on $\mathbb{R}^2$: kernel and image are both $\operatorname{span}\{e_1\}$.
(D) True Theorem 8.28(b).
Answer (A), (B) and (D).

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.

(A) True $\frac{d}{dx}\int_0^xp=p$ (fundamental theorem of calculus).
(B) False $JD(p)=\int_0^xp'=p-p(0)$; e.g. $JD(1)=0$.
(C) True $D(Jp)=p$ shows $D$ is onto; $D(1)=0$ shows it is not one-to-one.
(D) True $Jp=0$ gives $p=DJp=0$; and every $Jp$ vanishes at $0$, so $1\notin\operatorname{im}J$.
Answer (A), (C) and (D). (In finite dimension (A) would force (B) — Theorem 8.21(d).)

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 $A$.

(A) True $\operatorname{tr}(P^{-1}AP)=\operatorname{tr}(APP^{-1})$.
(B) True $B^{T}=(P^{-1}AP)^{T}=P^{T}A^{T}(P^{T})^{-1}$.
(C) False $A=\operatorname{diag}(1,2)$ and $B=\begin{pmatrix}1&1\\0&2\end{pmatrix}$ are similar ($P^{-1}BP=A$ for $P=\begin{pmatrix}1&1\\0&1\end{pmatrix}$), but $AB=\begin{pmatrix}1&1\\0&4\end{pmatrix}\ne\begin{pmatrix}1&2\\0&4\end{pmatrix}=BA$.
(D) False $I_2$ and $\begin{pmatrix}1&1\\0&1\end{pmatrix}$: trace $2$, determinant $1$, rank $2$, not similar.
Answer (A) and (B).
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min (plus 30 min for duals and nilpotents if time allows).

  1. 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 $\mathcal{L}(V,W)$.
  2. 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–Teaching NexusResearch-tutored · optional, about 45 minutes

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.

  1. 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 $c=P^{T}s$.
  2. Compute $c$ for the signal $s=(10,12,11,9,4,2,3,5)$.
  3. 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.
  4. 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.

Instructor note: $c_0=56/\sqrt8\approx19.799$, $c_1=28/\sqrt8\approx9.900$, $c_2=1$, $c_3=-1$, and $c_4,\dots,c_7=-\sqrt2,\sqrt2,\sqrt2,-\sqrt2$. Keeping $c_0,c_1$ gives $\tilde s=(10.5,10.5,10.5,10.5,3.5,3.5,3.5,3.5)$ with error $\sqrt{10}\approx3.16$, about $14\%$ of $\|s\|=\sqrt{500}\approx22.36$. Keeping the two largest samples ($12$ and $11$) leaves error $\sqrt{235}\approx15.3$. Dropping only the two smallest Haar coordinates ($\pm1$) gives error $\sqrt2$. Because $P$ is orthogonal, the error equals the length of the discarded coordinate vector (Parseval, §10.1).

🧪 Self-check quiz

0/5 correct

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