I'm trying to prove the existence of the linear map $f^*$ in the following theorem about the adjoint:

Let $f: V \to W$ be a linear map, with $V$ finite-dimensional and $V,W$ inner product spaces over the same field. Then there is a unique linear map $f^*: W \to V$ such that $$\langle f(v),w \rangle = \langle v, f^*(w)\rangle$$ for all $v\in V, w \in W$.

Here's how far I got so far.

Try 1:

Fix $w \in W$. Because $V$ is finite-dimensional there exists a unique $x \in V$ for all linear forms $\psi_1 = \langle f(\cdot),w \rangle \in V^*$ --- namely we have the isomorphism $V \to V^*$. Since the linear form $\langle f(\cdot),w \rangle$ is uniquely determined by $w \in W$, define the linear map $f^*$ by $x = f^*(w)$. Let $\psi_2 = \langle \cdot,f^*(w) \rangle \in V^*$ and $\theta = \psi_1 - \psi_2 \in V^*$.

Now consider ker$(\theta)$. If ker$(\theta)$ = $V$ then our claim is proven --- namely $\psi_1 = \psi_2$.

Mistake: Firstly $\psi_1 \in W^*$, and not in $V^*$. And because $W \to W^*$ need not be an isomorphism the argument doens't follow.

EDIT: $\psi_1 \in V^*$ indeed. Namely $\psi_1: v \mapsto \langle f(v),w \rangle$ and $w \in W$ is fixed. However note that $\psi_1$ uses an inner product on $W$.

EDIT: Note that rk$(\theta) = 1$, so we have dim$\big($ker$(\theta)\big) = $ dim$(V) - 1$. Which fails my attempted proof. Namely take dim$(V) = 1$. So that, ker$(\theta)$ = $\{0\}$. And thus for all $0 \neq v \in V$ we have $\psi_1 \neq \psi_2$.

Try 2: the same as Try 1 but with $v = x$.

I'm assuming that the argument can be proven by a less cumbersome method. Also I think that my method is not correct, since I seem to get stuck.

Suggestions or an alternative method are welcome.

EDIT: Here's a summary of Fischer's proof.

Define the linear map $\lambda: W \to V^*$ by $$\lambda(w) = \langle \cdot, w \rangle_W \circ f.$$ Next define the isomorphism $\sigma: V \to V^*$ by $$\sigma(v_0) = \langle \cdot, v_0 \rangle_V.$$ Now, consider the composition $$\varphi: \sigma^{-1} \circ \lambda: W \to V.$$ We then have $$\langle v, \varphi(w) \rangle_V = \langle f(v), w \rangle_W$$ for all $v \in V, w \in W$. We can now define $f^*:= \varphi$. This proves the theorem for the existence part.

For uniqueness, assume that $\psi$ satisfies $\langle f(v),w \rangle_W = \langle v, \psi(w) \rangle_V$. Fixing $w_0 \in W$, we get $$\langle v,\psi(w_0) - f^*(w_0) \rangle_V = 0$$ for all $v \in V$. Choosing $v = \psi(w_0) - f^*(w_0)$ we get $\psi = f^*$, which proves the uniqueness part.

  • 1
    $\begingroup$ Your notation seems to imply that both spaces are endowed with the same inner product, is this the case? As a hint may I suggest that you write out the statement you wish $f^*$ to satisfy, and then see if you can construct $f^*$ using the function $f$ and the isomorphisms between $W,W^*$ and $V,V^*$. $\endgroup$
    – R R
    Dec 3, 2013 at 19:03
  • $\begingroup$ $V$ and $W$ are not endowed with the same inner product and I have listed the mistake in the question. $\endgroup$ Dec 3, 2013 at 20:13
  • $\begingroup$ What do you mean with the isomorphisms between $W,W^*$ and $V,V^*$? If you mean Hom$(W,W^*) \to$Hom$(V,V^*)$, then why does there exist such an isomorphim. Note that the isomorphism $W \to W^*$ need not exist. $\endgroup$ Dec 4, 2013 at 1:30

3 Answers 3


Let's take a step back, and look at it from a more abstract perspective.

First, we define a map $\lambda \colon W \to V^\ast$ by

$$\lambda(w) \colon v \mapsto \langle f(v), w\rangle_W,$$

or $\lambda(w) = \langle\,\cdot\,,w\rangle_W \circ f$. Now, looking at it, we notice that $\lambda$ is linear (or, if the scalar field is $\mathbb{C}$, and the inner product linear in the first and antilinear in the second argument, $\lambda$ is antilinear; if the spaces are complex and the inner product is antilinear in the first, and linear in the second argument, consider $\langle w,f(v)\rangle_W$ instead):

$$\begin{align} \lambda(w_1+w_2)(v) &= \langle f(v), w_1 + w_2\rangle_W\\ &= \langle f(v), w_1\rangle_W + \langle f(v),w_2\rangle_W\\ &= \lambda(w_1)(v) + \lambda(w_2)(v)\\ &= \bigl(\lambda(w_1) + \lambda(w_2)\bigr)(v), \end{align}$$

and since $v$ was arbitrary, that means $\lambda(w_1+w_2) = \lambda(w_1) + \lambda(w_2)$. And for a scalar $s$, we have

$$\begin{align} \lambda(s\cdot w)(v) &= \langle f(v), s\cdot w\rangle_W\\ &= s\cdot \langle f(v), w\rangle_W\\ &= s\cdot \lambda(w)(v)\\ &= \bigl(s\cdot \lambda(w)\bigr)(v), \end{align}$$

that is, $\lambda(s\cdot w) = s\cdot \lambda(w)$ (for complex scalars, we would have to conjugate $s$ and $\lambda$ would be antilinear).

Now, with the (anti-) isomorphism $\sigma \colon V \to V^\ast$, given by $\sigma(v) = \langle\,\cdot\,,v\rangle_V$, i.e. $\sigma(v) \colon w \mapsto \langle w, v\rangle_V$ (or $\sigma(v) = \langle v,\,\cdot\,\rangle_V$, i.e. $\sigma(v) \colon w \mapsto \langle v,w\rangle_V$ if $V$ is a complex vector space and the inner product is antilinear in the first argument, as is customary in physics), consider the composition

$$\varphi = \sigma^{-1} \circ \lambda \colon W \to V.$$

Then $\varphi$ is linear (as the composition of two linear maps in the real case, and as the composition of two antilinear maps in the complex case), and for $v\in V,\; w \in W$, we have

$$\begin{align} \langle v, \varphi(w)\rangle_V &= \langle v, \sigma^{-1}(\lambda(w))\rangle_V\\ &= \lambda(w)(v)\tag{definition of $\sigma$}\\ &= \langle f(v),w\rangle_W,\tag{definition of $\lambda$} \end{align}$$

and that means we can define $f^\ast := \varphi = \sigma^{-1}\circ \lambda$.

It remains to be seen that there is only one such map $W \to V$.

Suppose that $\psi \colon W \to V$ is a map with $\langle f(v), w\rangle_W = \langle v, \psi(w)\rangle_V$ for all $v\in V,\; w\in W$.

Fix an arbitrary $w_0\in W$. Then for all $v\in V$

$$\begin{align} \langle v,\psi(w_0) - f^\ast(w_0)\rangle_V &= \langle v,\psi(w_0)\rangle_V - \langle v, f^\ast(w_0)\rangle_V\\ &= \langle f(v), w_0\rangle_W - \langle f(v), w_0\rangle_W\\ &= 0. \end{align}$$

In particular, we can choose $v = \psi(w_0) - f^\ast(w_0)$, so we have

$$\langle \psi(w_0) - f^\ast(w_0), \psi(w_0) - f^\ast(w_0)\rangle_V = 0,$$

which means $\psi(w_0) - f^\ast(w_0) = 0$. Since $w_0 \in W$ was arbitrary, it follows that $\psi = f^\ast$, and the uniqueness is also established.

  • $\begingroup$ How does $\langle v, \sigma^{-1}(\lambda (w)) \rangle_V = \lambda (w)(v)$ follow from the definition of $\sigma$? $\endgroup$ Dec 3, 2013 at 23:20
  • 3
    $\begingroup$ The definition of $\sigma \colon V \to V^\ast$ is $\sigma(v_0)(v) = \langle v, v_0\rangle_V$. And that means $\mu(v) = \sigma(\sigma^{-1}(\mu))(v) = \langle v, \sigma^{-1}(\mu)\rangle_V$ for $\mu \in V^\ast$. $\endgroup$ Dec 3, 2013 at 23:24
  • $\begingroup$ How is $\langle \cdot, w \rangle_W \circ f$ a correct notation for $\lambda(w)$? $\endgroup$ Dec 3, 2013 at 23:56
  • 1
    $\begingroup$ Both denote the map $v \mapsto \langle f(v), w\rangle_W$. Of course one may have qualms about a notation $\langle \,\cdot\,,w\rangle_W$, then ignore that part. $\endgroup$ Dec 3, 2013 at 23:59
  • 2
    $\begingroup$ $\langle v,w\rangle_W$ doesn't even make sense, since $v\in V$. It's the composition of the two maps $f\colon v \mapsto f(v)$, and $\langle\,\cdot\,,w\rangle_W\colon z \mapsto \langle z, w\rangle_W$. The value that $z$ is instantiated with is the result of the application of $f$, namely $f(v)$, so $$\bigl(\langle\,\cdot\,,w\rangle_W\circ f\bigr)(v) = \langle\,\cdot\,,w\rangle_W \bigl(f(v)\bigr) = \langle f(v),w\rangle_W.$$ $\endgroup$ Dec 4, 2013 at 0:09

Fix $w\in W$. Consider the linear form on $V$ that maps $v\in V$ to $\langle f(v),w\rangle$. Let $f^{*}w$ be the unique vector in V such that this linear form is given by taking inner products with $f^{*}w$. So, the $f^{*}w$ is the unique vector such that $\langle f(v),w\rangle= \langle v, f^{*} w\rangle \; \forall v\in V.$

I think the existence of the adjoint (as described above) depends a lot on the fact that

$f^{*}w$ is the unique vector in V such that this linear form is given by taking inner products with $f^{*}w$

which you seem to also use in your fist try. Does this help?

  • $\begingroup$ It doesn't answer my question: I want to prove that such a $f^*$ exists. Indeed, it depends on equality of the two linear forms, but how to show that such a $f^*$ exists? $\endgroup$ Dec 2, 2013 at 17:21

This essentially follows from Riesz Representation theorem. That is, each linear functional $f(x)$ may be written as $<x,w_f>$ for some vector $w_f$. To prove Riesz Representation theorem, fix an orthonormal basis and write $x$ as linear combination of those...


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.