books/napkin/dual-trace.tex

\chapter{Dual space and trace}
You may have learned in high school that given a matrix
\[
	\begin{bmatrix}
		a & c \\
		b & d
	\end{bmatrix}
\]
the trace is the sum along the diagonals $a+d$
and the determinant is $ad-bc$.
But we know that a matrix is somehow
just encoding a linear map using a choice of basis.
Why would these random formulas somehow not
depend on the choice of a basis?

In this chapter, we are going to
give an intrinsic definition of $\Tr T$,
where $T \colon V \to V$ and $\dim V < \infty$.
This will give a coordinate-free definition
which will in particular imply the trace $a+d$
doesn't change if we take a different basis.

In doing so, we will introduce two new constructions:
the \emph{tensor product} $V \otimes W$
(which is a sort of product of two spaces,
with dimension $\dim V \cdot \dim W$)
and the \emph{dual space} $V^\vee$,
which is the set of linear maps $V \to k$ (a $k$-vector space).
Later on, when we upgrade from a vector space $V$
to an inner product space,
we will see that the dual space $V^\vee$ gives a nice
interpretation of the ``transpose'' of a matrix.
You'll already see some of that come through here.

The trace is only defined for finite-dimensional
vector spaces, so if you want you can restrict
your attention to finite-dimensional vector spaces for this chapter.
(On the other hand we do not need the
ground field to be algebraically closed.)

The next chapter will then do the same for the determinant.

\section{Tensor product}
\prototype{$\RR[x] \otimes \RR[y] = \RR[x,y]$.}
We know that $\dim (V \oplus W) = \dim V + \dim W$,
even though as sets $V \oplus W$ looks like $V \times W$.
What if we wanted a real ``product'' of spaces,
with multiplication of dimensions?

For example, let's pull out
my favorite example of a real vector space, namely
\[ V = \left\{ ax^2 + bx + c \mid a,b,c \in \RR \right\}. \]
Here's another space, a little smaller:
\[ W = \left\{ dy + e \mid d,e \in \RR \right\}. \]
If we take the direct sum, then we would get some rather unnatural
vector space of dimension five
(whose elements can be thought of as pairs $(ax^2+bx+c,dy+e)$).
But suppose we want a vector space
whose elements are \emph{products} of polynomials in $V$ and $W$;
it would contain elements like $4x^2y + 5xy + y + 3$.
In particular, the basis would be
\[ \left\{ x^2y, x^2, xy, x, y, 1 \right\} \]
and thus have dimension six.

For this we resort to the \emph{tensor product}.
It does exactly this, except that the ``multiplication''
is done by a scary\footnote{%
	Seriously, $\otimes$ looks \emph{terrifying} to non-mathematicians,
	and even to many math undergraduates.}
symbol $\otimes$:
think of it as a ``wall'' that separates the elements
between the two vector spaces.
For example, the above example might be written as
\[ 4x^2 \otimes y + 5x \otimes y + 1 \otimes y + 3 \otimes 1. \]
(This should be read as $(4x^2 \otimes y) + (5x \otimes y) + \dots$;
addition comes after $\otimes$.)
Of course there should be no distinction
between writing $4x^2 \otimes y$ and $x^2 \otimes 4y$
or even $2x^2 \otimes 2y$.
While we want to keep the $x$ and $y$ separate,
the scalars should be free to float around.

Of course, there's no need to do everything
in terms of just the monomials.
We are free to write
\[ (x + 1) \otimes (y + 1). \]
If you like, you can expand this as
\[ x \otimes y + 1 \otimes y + x \otimes 1 + 1 \otimes 1. \]
Same thing.
The point is that we can take any two of our polynomials
and artificially ``tensor'' them together.

The definition of the tensor product does exactly this,
and nothing else.\footnote{I'll only define this
	for vector spaces for simplicity.
	The definition for modules over a commutative ring $R$ is exactly the same.}
\begin{definition}
	Let $V$ and $W$ be vector spaces over the same field $k$.
	The \vocab{tensor product} $V \otimes_k W$ is the abelian group
	generated by elements of the form $v \otimes w$, subject to relations
	\begin{align*}
		(v_1 + v_2) \otimes w &= v_1 \otimes w + v_2 \otimes w \\
		v \otimes (w_1 + w_2) &= v \otimes w_1 + v \otimes w_2 \\
		(c \cdot v) \otimes w &= v \otimes (c \cdot w).
	\end{align*}
	As a vector space,
	its action is given by
	$c \cdot (v \otimes w) = (c \cdot v) \otimes w = v \otimes (c \cdot w)$.
\end{definition}
Here's another way to phrase the same idea.
We define a \vocab{pure tensor} as an
element of the form $v \otimes w$ for $v \in V$ and $w \in W$.
But we let the $\otimes$ wall be ``permeable'' in the sense that
\[ (c \cdot v) \otimes w = v \otimes (c \cdot w) = c \cdot (v \otimes w) \]
and we let multiplication and addition distribute as we expect.
Then $V \otimes W$ consists of sums of pure tensors.

\begin{example}
	[Infinite-dimensional example of tensor product: two-variable polynomials]
	Although it's not relevant to this chapter,
	this definition works equally well with infinite-dimensional
	vector spaces.
	The best example might be
	\[ \RR[x] \otimes_\RR \RR[y] = \RR[x,y]. \]
	That is, the tensor product of polynomials in $x$
	with real polynomials in $y$
	turns out to just be two-variable polynomials $\RR[x,y]$.
\end{example}

\begin{remark}
	[Warning on sums of pure tensors]
	Remember the elements of $V \otimes_k W$
	really are \emph{sums} of these pure tensors!
	If you liked the previous example,
	this fact has a nice interpretation ---
	not every polynomial in $\RR[x,y] = \RR[x] \otimes_\RR \RR[y]$ factors
	as a polynomial in $x$ times a polynomial in $y$
	(i.e.\ as pure tensors $f(x) \otimes g(y)$).
	But they all can be written as sums of pure tensors $x^a \otimes y^b$.
\end{remark}


As the example we gave suggested,
the basis of $V \otimes_k W$ is literally the ``product''
of the bases of $V$ and $W$.
In particular, this fulfills our desire that
$\dim (V \otimes_k W) = \dim V \cdot \dim W$.
\begin{proposition}[Basis of $V \otimes W$]
	Let $V$ and $W$ be finite-dimensional $k$-vector spaces.
	If $e_1, \dots, e_m$ is a basis of $V$ and $f_1, \dots, f_n$ is a basis of $W$,
	then the basis of $V \otimes_k W$
	is precisely $e_i \otimes f_j$, where $i=1,\dots,m$ and $j=1,\dots,n$.
\end{proposition}
\begin{proof}
	Omitted; it's easy at least to see that this basis is spanning.
\end{proof}

\begin{example}[Explicit computation]
	Let $V$ have basis $e_1$, $e_2$ and $W$ have basis $f_1, f_2$.
	Let $v = 3e_1 + 4e_2 \in V$ and $w = 5f_1 + 6f_2 \in W$.
	Let's write $v \otimes w$ in this basis for $V \otimes_k W$:
	\begin{align*}
		v \otimes w &= (3e_1+4e_2) \otimes (5f_1+6f_2) \\
		&= (3e_1) \otimes (5f_1) +  (4e_2) \otimes (5f_1)
		+ (3e_1) \otimes (6f_2) + (4e_2) \otimes (6f_2) \\
		&= 15 (e_1 \otimes f_1) + 20(e_2 \otimes f_1)
		+ 18 (e_1 \otimes f_2) + 24(e_2 \otimes f_2).
	\end{align*}
	So you can see why tensor products are a nice ``product'' to
	consider if we're really interested in $V \times W$
	in a way that's more intimate than just a direct sum.
\end{example}

\begin{abuse}
	Moving forward, we'll almost always abbreviate $\otimes_k$ to just $\otimes$,
	since $k$ is usually clear.
\end{abuse}
\begin{remark}
	Observe that to define a linear map $V \otimes W \to X$,
	I only have to say what happens to each pure tensor $v \otimes w$,
	since the pure tensors \emph{generate} $V \otimes W$.
	But again, keep in mind that
	$V \otimes W$ consists of \emph{sums} of these pure tensors!
	In other words, $V \otimes W$ is generated by pure tensors.
\end{remark}

\begin{remark}
	Much like the Cartesian product $A \times B$ of sets,
	you can tensor together any two vector spaces $V$ and $W$ over the same field $k$;
	the relationship between $V$ and $W$ is completely irrelevant.
	One can think of the $\otimes$ as a ``wall'' through which one can pass
	scalars in $k$, but otherwise keeps the elements of $V$ and $W$ separated.
	Thus, $\otimes$ is \textbf{content-agnostic}.

	This also means that even if $V$ and $W$ have some relation to each other,
	the tensor product doesn't remember this.
	So for example $v \otimes 1 \neq 1 \otimes v$,
	just like $(g,1_G) \neq (1_G,g)$ in the group $G \times G$.
\end{remark}


\section{Dual space}
\prototype{Rotate a column matrix by $90$ degrees.}

Consider the following vector space:
\begin{example}
	[Functions from $\RR^3 \to \RR$]
	The set of real functions $f(x,y,z)$ is an
	infinite-dimensional real vector space.
	Indeed, we can add two functions to get $f+g$,
	and we can think of functions like $2f$.
\end{example}
This is a terrifyingly large vector space,
but you can do some reasonable reductions.
For example, you can restrict your attention to just
the \emph{linear maps} from $\RR^3$ to $\RR$.

That's exactly what we're about to do.
This definition might seem strange at first, but bear with me.

\begin{definition}
	Let $V$ be a $k$-vector space.
	Then $V^\vee$, the \vocab{dual space} of $V$, is defined
	as the vector space whose elements are \emph{linear maps from $V$ to $k$}.
\end{definition}
The addition and multiplication are pointwise:
it's the same notation we use when we write $cf+g$ to mean $c \cdot f(x) + g(x)$.
The dual space itself is less easy to think about.

Let's try to find a basis for $V^\vee$.
First, here is a very concrete interpretation of the vector space.
Suppose for example $V = \RR^3$.
We can think of elements of $V$ as column matrices, like
\[ v = \begin{bmatrix}
		2 \\ 5 \\ 9
	\end{bmatrix}
	\in V. \]
Then a linear map $f \colon V \to k$ can be interpreted as a \emph{row matrix}:
\[
	f = \begin{bmatrix}
		3 & 4 & 5
	\end{bmatrix}
	\in V^\vee. \]
Then
\[
	f(v) = \begin{bmatrix}
		3 & 4 & 5
	\end{bmatrix}
	\begin{bmatrix}
		2 \\ 5 \\ 9
	\end{bmatrix}
	= 71. \]

More precisely: \textbf{to specify a linear map $V \to k$,
I only have to tell you where each basis element of $V$ goes}.
In the above example, $f$ sends $e_1$ to $3$, $e_2$ to $4$, and $e_3$ to $5$.
So $f$ sends \[ 2e_1 + 5e_2 + 9e_3 \mapsto 2 \cdot 3 + 5 \cdot 4 + 9 \cdot 5 = 71. \]

Let's make all this precise.
\begin{proposition}[The dual basis for $V^\vee$]
	Let $V$ be a finite-dimensional vector space with basis $e_1, \dots, e_n$.
	For each $i$ consider the function $e_i^\vee \colon V \to k$
	defined by
	\[
		e_i^\vee(e_j)
		= \begin{cases}
			1 & i=j \\
			0 & i \neq j.
		\end{cases}
	\]
	In more humane terms, $e_i^\vee(v)$
	gives the coefficient of $e_i$ in $v$.

	Then $e_1^\vee$, $e_2^\vee$, \dots, $e_n^\vee$ is a basis of $V^\vee$.
\end{proposition}

\begin{example}[Explicit example of element in $V^\vee$]
	In this notation, $f = 3e_1^\vee + 4e_2^\vee + 5e_3^\vee$.
	Do you see why the ``sum'' notation works as expected here?
	Indeed
	\begin{align*}
		f(e_1) &= (3e_1^\vee + 4e_2^\vee + 5e_3^\vee)(e_1) \\
		&= 3e_1^\vee(e_1) + 4e_2^\vee(e_1) + 5e_3^\vee(e_1) \\
		&= 3 \cdot 1 + 4 \cdot 0 + 5 \cdot 0 = 3.
	\end{align*}
	That's exactly what we wanted.
\end{example}

You might be inclined to point out that $V \cong V^\vee$ at this point,
with an isomorphism given by $e_i \mapsto e_i^\vee$.
You might call it ``rotating the column matrix by $90\dg$''.

This statement is technically true,
but for a generic vector space $V$ without any extra information,
you can just think of this as an artifact of the $\dim V = \dim V^\vee$
(as \emph{any} two vector spaces of equal dimension are isomorphic).
Most importantly, the isomorphism given above depends
on what basis you picked.

\begin{remark*}
	[Explicit example showing that the isomorphism $V \to V^\vee$
	given above is unnatural]
	Alice or Bob are looking at the same two-dimensional real vector space
	\[ V = \left\{ (x,y,z) \mid x+y+z = 0 \right\}. \]
	Also, let $v_{\text{example}} = (3,5,-8)$
	be an example of an arbitrary element of $V$ for concreteness.

	Suppose Alice chooses the following basis vectors for $V$.
	\begin{align*}
		e_1 &= (1,0,-1) \\
		e_2 &= (0,1,-1).
	\end{align*}
	Alice uses this to construct an isomorphism $A \colon V \to V^\vee$
	as described above, and considers $e_1^\vee = A(e_1^\vee)$.
	The element $e_1^\vee \in V^\vee$ is a
	function $e_1^\vee \colon V \to \RR$,
	meaning Alice can plug any vector in $V$ into it.
	As an example, for $v_{\text{example}}$
	\[ e_1^\vee(v_{\text{example}})
		= e_1\left( (3,5,-8) \right)
		= e_1^\vee\left( 3e_1 - 5e_2 \right) = 3. \]
	Meanwhile, Bob chooses the different basis vectors
	\begin{align*}
		f_1 &= (1,0,-1) \\
		f_2 &= (1,-1,0).
	\end{align*}
	This gives Bob an isomorphism $B \colon V \to V^\vee$,
	and a corresponding $f_1^\vee = B(f_1)$.
	Bob can also evaluate it anywhere, e.g.
	\[ f_1^\vee\left( v_{\text{example}} \right)
		f_1^\vee\left( (3, 5, -8) \right)
		= f_1^\vee\left( 8f_1 - 5f_2 \right) = 8. \]

	It follows that $e_1^\vee = A\left( (1,0,-1) \right)$
	and $f_1^\vee = B\left( (1,0,-1) \right)$
	are different elements of $V^\vee$.
	In other words Alice and Bob got different isomorphisms
	because they picked different bases.
\end{remark*}

\section{$V^\vee \otimes W$ gives matrices from $V$ to $W$}
Goal of this section:
\begin{moral}
	If $V$ and $W$ are finite-dimensional $k$-vector spaces
	then $V^\vee \otimes W$ represents linear maps $V \to W$.
\end{moral}

Here's the intuition.
If $V$ is three-dimensional and $W$ is five-dimensional, then we can think
of the maps $V \to W$ as a $5 \times 3$ array of numbers.
We want to think of these maps as a vector space:
(since one can add or scale matrices).
So it had better be a vector space with dimension $15$,
but just saying ``$k^{\oplus 15}$'' is not really that satisfying
(what is the basis?).

To do better, we consider the tensor product
\[ V^\vee \otimes W \]
which somehow is a product of maps out of $V$ and the target space $W$.
We claim that this is in fact the space we want:
i.e.\ \textbf{there is a natural bijection between elements of $V^\vee \otimes W$
and linear maps from $V$ to $W$}.

First, how do we interpret an element of $V^\vee \otimes W$ as a map $V \to W$?
For concreteness, suppose $V$ has a basis $e_1$, $e_2$, $e_3$,
and $W$ has a basis $f_1$, $f_2$, $f_3$, $f_4$, $f_5$.
Consider an element of $V^\vee \otimes W$, say
\[ e_1^\vee \otimes (f_2 + 2f_4) + 4e_2^\vee \otimes f_5. \]
We want to interpret this element as a function $V \to W$:
so given a $v \in V$,
we want to output an element of $W$.
There's really only one way to do this:
feed in $v \in V$ into the $V^\vee$ guys on the left.
That is, take the map
\[ v \mapsto e_1^\vee(v) \cdot (f_2 + 2f_4) + 4e_2^\vee(v) \cdot f_5 \in W. \]
So, there's a natural way to interpret any element
$\xi_1 \otimes w_1 + \dots + \xi_m \otimes w_m \in V^\vee \otimes W$
as a linear map $V \to W$.
The claim is that in fact, every linear map $V \to W$ has
such an interpretation.

First, for notational convenience,
\begin{definition}
	Let $\Hom(V,W)$ denote the set of linear maps from $V$ to $W$
	(which one can interpret as matrices which send $V$ to $W$),
	viewed as a vector space over $k$.
	(The ``$\Hom$'' stands for homomorphism.)
\end{definition}
\begin{ques}
	Identify $\Hom(V,k)$ by name.
\end{ques}

We can now write down something that's more true generally.
\begin{theorem}[$V^\vee \otimes W$ $\iff$ linear maps $V \to W$]
	\label{thm:vect_hom_dualization}
	Let $V$ and $W$ be finite-dimensional vector spaces.
	We described a map
	\[ \Psi \colon V^\vee \otimes W \to \Hom(V,W) \]
	by sending $\xi_1 \otimes w_1 + \dots + \xi_m \otimes w_m$ to the linear map
	\[ v \mapsto \xi_1(v) w_1 + \dots + \xi_m(v) w_m. \]
	Then $\Psi$ is an isomorphism of vector spaces, i.e.\ every linear map $V \to W$
	can be uniquely represented as an element of $V^\vee \otimes W$ in this way.
\end{theorem}

The above is perhaps a bit dense, so here is a concrete example.
\begin{example}[Explicit example]
	Let $V = \RR^2$ and take a basis $e_1$, $e_2$ of $V$.
	Then define $T : V \to V$ by
	\[ T = \begin{bmatrix}
			1 & 2 \\ 3 & 4
		\end{bmatrix}. \]
	Then we have
	\[ \Psi(e_1^\vee \otimes e_1 + 2e_2^\vee \otimes e_1
		+ 3e_1^\vee \otimes e_2 + 4e_2^\vee \otimes e_2) = T. \]
	The beauty is that the $\Psi$ definition is basis-free;
	thus even if we change the basis,
	although the above expression will look completely different,
	the \emph{actual element} in $V^\vee \otimes V$ doesn't change.
\end{example}

Despite this, we'll indulge ourselves in using coordinates for the proof.
\begin{proof}
	[Proof of \Cref{thm:vect_hom_dualization}]
	This looks intimidating, but it's actually not difficult.
	We proceed in two steps:
	\begin{enumerate}
		\ii First, we check that $\Psi$ is \emph{surjective};
		every linear map has at least one representation in $V^\vee \otimes W$.
		To see this, take any $T : V \to W$.
		Suppose $V$ has basis $e_1$, $e_2$, $e_3$ and that
		$T(e_1) = w_1$, $T(e_2) = w_2$ and $T(e_3) = w_3$.
		Then the element
		\[ e_1^\vee \otimes w_1 + e_2^\vee \otimes w_2 + e_3^\vee \otimes w_3 \]
		works, as it is contrived to agree with $T$ on the basis elements $e_i$.
		\ii So it suffices to check now that $\dim V^\vee \otimes W = \dim \Hom(V,W)$.
		Certainly, $V^\vee \otimes W$ has dimension $\dim V \cdot \dim W$.
		But by viewing $\Hom(V,W)$ as $\dim V \cdot \dim W$ matrices, we see that
		it too has dimension $\dim V \cdot \dim W$. \qedhere
	\end{enumerate}
\end{proof}
So there is a \textbf{natural isomorphism} $V^\vee \otimes W \cong \Hom(V,W)$.
While we did use a basis liberally in the
\emph{proof that it works}, this doesn't change the
fact that the isomorphism is ``God-given'',
depending only on the spirit of $V$ and $W$ itself
and not which basis we choose to express the vector spaces in.


\section{The trace}
We are now ready to give the definition of a trace.
Recall that a square matrix $T$ can be thought of as a map $T \colon V \to V$.
According to the above theorem,
\[ \Hom(V, V) \cong V^\vee \otimes V \]
so every map $V \to V$ can be thought of as an element of $V^\vee \otimes V$.
But we can also define an
\emph{evaluation map} $\opname{ev} : V^\vee \otimes V \to k$
by ``collapsing'' each pure tensor: $f \otimes v \mapsto f(v)$.
So this gives us a composed map
\begin{center}
\begin{tikzcd}
	\Hom(V, V) \ar[r, "\cong"] & V^\vee \otimes V \ar[r, "\opname{ev}"] & k.
\end{tikzcd}
\end{center}
This result is called the \vocab{trace} of a matrix $T$.

\begin{example}[Example of a trace]
	Continuing the previous example,
	\[ \Tr T = e_1^\vee(e_1) + 2e_2^\vee(e_1)
		+ 3e_1^\vee(e_2) + 4e_2^\vee(e_2)
		= 1 + 0 + 0 + 4 = 5. \]
	And that is why the trace is the sum of the diagonal entries.
\end{example}

\section{\problemhead}

\begin{problem}
	[Trace is sum of eigenvalues]
	Let $V$ be an $n$-dimensional vector space
	over an algebraically closed field $k$.
	Let $T \colon V \to V$ be a linear map with
	eigenvalues $\lambda_1$, $\lambda_2$, \dots, $\lambda_n$
	(counted with algebraic multiplicity).
	Show that $\Tr T = \lambda_1 + \dots + \lambda_n$.
	\begin{hint}
		Follows by writing $T$ in an eigenbasis:
		then the diagonal entries are the eigenvalues.
	\end{hint}
\end{problem}

\begin{dproblem}
	[Product of traces]
	Let $T \colon V \to V$ and $S \colon W \to W$ be linear maps
	of finite-dimensional vector spaces $V$ and $W$.
	Define $T \otimes S \colon V \otimes W \to V \otimes W$
	by $v \otimes w \mapsto T(v) \otimes S(w)$.
	Prove that \[ \Tr(T \otimes S) = \Tr(T) \Tr(S). \]
	\begin{hint}
		Again one can just take a basis.
	\end{hint}
\end{dproblem}

\begin{dproblem}
	[Traces kind of commute]
	\gim
	Let $T \colon V \to W$ and $S \colon W \to V$ be linear maps
	between finite-dimensional vector spaces $V$ and $W$.
	Show that \[ \Tr(T \circ S) = \Tr(S \circ T). \]
	\begin{hint}
		One solution is to just take a basis.
		Otherwise, interpret $T \otimes S \mapsto \Tr(T \circ S)$ as a
		linear map $(V^\vee \otimes W) \otimes (W^\vee \otimes V) \to k$,
		and verify that it is commutative.
	\end{hint}
	\begin{sol}
		Although we could give a coordinate calculation,
		we instead opt to give a cleaner proof.
		This amounts to drawing the diagram
		\begin{center}
			\begin{tikzcd}[column sep=tiny]
			& (W^\vee \otimes V) \otimes (V^\vee \otimes W) \ar[d]
				\ar[r, equals] \ar[ld, "\text{compose}"']
				& (V^\vee \otimes W) \otimes (W^\vee \otimes V) \ar[d]
				\ar[rd, "\text{compose}"]
				\\
			\Hom(W, W) \ar[r, leftrightarrow] \ar[rd, "\Tr"']
				& W^\vee \otimes W \ar[d, "\opname{ev}"]
				& V^\vee \otimes V \ar[r, leftrightarrow] \ar[d, "\opname{ev}"]
				& \Hom(V, V) \ar[ld, "\Tr"']
				\\
			& k \ar[r, equals] & k
		\end{tikzcd}
		\end{center}
		It is easy to check that the center rectangle commutes,
		by checking it on pure tensors $\xi_W \otimes v \otimes \xi_V \otimes w$.
		So the outer hexagon commutes and we're done.
		This is really the same as the proof with bases;
		what it amounts to is checking the assertion is true for
		matrices that have a $1$ somewhere and $0$ elsewhere,
		then extending by linearity.
	\end{sol}
\end{dproblem}

\begin{problem}
	[Putnam 1988]
	\gim
	Let $V$ be an $n$-dimensional vector space.
	Let $T : V \to V$ be a linear map and suppose there exists $n+1$ eigenvectors,
	any $n$ of which are linearly independent.
	Does it follow that $T$ is a scalar multiple of the identity?
	\begin{hint}
		Look at the trace of $T$.
	\end{hint}
	\begin{sol}
		See \url{https://mks.mff.cuni.cz/kalva/putnam/psoln/psol886.html}.
	\end{sol}
\end{problem}