The fact -- as is often introduced in an introductory general relativity or tensor calculus course -- that the gradient is a
covector seems rather bizarre to someone who's always seen the gradient as the "steepest ascent vector". Surely, the direction of steepest ascent is, you know, a direction -- an arrow. And what even is a covector, anyway?
Let's think about differentiating with respect to vectors. The idea we have is that $\frac{\partial f}{\partial \vec x}$ needs to contain all the information -- each of the $\frac{\partial f}{\partial x_i}$. And analogously for derivatives with respect to tensors. You might think we could just create an array with the same dimensions containing each derivative -- much like the gradient, Hessian, etc. that we're used to -- i.e.
$$\nabla f=\left[ {{\partial ^i}f} \right]$$
$$\nabla^2f = \left[ {{\partial ^i}{\partial ^j}f} \right]$$
(I'm using $\nabla^2$ for the Hessian -- and will do so in the rest of the article -- but it's too widely used for its trace the Laplacian, which should be represented as $|\nabla|^2$) etc. But you might get the sense that this feels just fundamentally wrong -- like you're giving the "division by tensor" object the structure of the same tensor, but you should somehow be giving it an "inverse" structure.
We want to construct a situation to see that the idea above -- of making the gradient ("derivative with respect to a vector") and Hessian ("derivative with respect to a rank-2 tensor") -- a vector and a rank-tensor doesn't work. We know such a situation can arrive when we have multiplication between the gradient and a vector, or the Hessian and a rank-2 tensor. For instance, for linear $f$:
$$f(\vec{x})-f(0)=\vec{x}\cdot\nabla f$$
But this is
wrong -- for any non-Euclidean manifold. For instance, if the metric tensor is something like $\rm{diag}(-1,1)$, this dot product gives:
$$f(\vec{x})-f(0)= - x\frac{{\partial f}}{{\partial x}} + y\frac{{\partial f}}{{\partial y}}$$
Which is just wrong. So instead, the gradient is a
covector, which we represent in Einstein notation using subscripts instead of superscripts:
$$f(\vec x) - f(0) = {x^i}{\partial _i}f$$
(As you can see, I omitted Einstein notation when I was writing the wrong equations -- seeing repeated indices on the same vertical alignment is physically painful.) If we want the
vector gradient -- for direction of steepest ascent or whatever -- you need to multiply by the metric tensor.
This also motivates the picture of seeing covectors as parallel surfaces whose normals are their vector versions -- in Euclidean geometry, it doesn't make a difference, but on a general setting, this normality is a bit weird. Think about this.
But I haven't really given a motivation for the metric tensor or how it comes up here -- for this, read on.
Let's talk about something completely different -- let's think about the derivative of functions from $\mathbb{C}\to\mathbb{R}$, $df/dz$. I don't know about you, but I like the complex numbers, and prefer them to $\mathbb{R}^2$, because pretty much anything I write with the complex numbers is well-defined, and easily so -- so I don't need to worry about whether $df/d \vec{x}$ makes any sense or not. Well, we can write:
$$\frac{df}{dz}=\frac{\partial f}{\partial x}\frac{\partial x}{\partial z}+\frac{\partial f}{\partial y}\frac{\partial y}{\partial z}\\\Rightarrow \frac{df}{dz}=\frac{\partial f}{\partial x}-\frac{\partial f}{\partial y}i$$
This $df/dz$ above is exactly the analog of the gradient for real-valued functions defined on the complex plane -- analogous to scalar multivariable functions.
What's the expression for the complex derivative of a complex function? Compute it -- it may look a bit different from the analogous tensor derivative -- think of traces and commutators.
Note: In actual complex calculus, complex differentiability is defined in a more restrictive way -- specifically one needs to satisfy the
Cauchy-Riemann equations, which makes the structure of complex functions fundamentally more special than that of multivariable functions, stuff like $dx/dz$ is even undefined, and the stuff we've written above isn't really relevant in complex analysis. It is, however, the "Wirtinger derivative".
Something interesting happened here, though -- we got a negative sign on the imaginary component of the derivative. The derivative got conjugated, or something -- and the reason this occurred is that $i^2=-1$ (so $1/i=-i$), and this leaves some sort of signature in our derivative.
Now let's (
non-rigorous alert!) think about how an analogous argument may be written for vectors.
$$\frac{{df}}{{d\vec x}} = \frac{{\partial f}}{{\partial x}}\frac{{\partial x}}{{\partial \vec x}} + \frac{{\partial f}}{{\partial y}}\frac{{\partial y}}{{\partial \vec x}}$$
What really is $\frac{{\partial x}}{{\partial \vec x}}$, though? We know that $\frac{\partial \vec x}{\partial x}=\vec{e_x}$. But what's the "inverse" of a vector? What does that even mean?
So we want to define some sort of a product, or multiplication, with vectors -- we want to define a thing that when multiplied by a vector gives a scalar. It sounds like we're talking about a dot product -- but the dot product lacks an important property we need to have division, it's not injective. I.e. $\vec{a}\cdot\vec{b}=c$ for fixed $\vec{a}$ and $c$ defines a whole plane of vectors $\vec{b}$, not a unique one. But if we added an additional component to our product, the cross product (or in more than three dimensions, the wedge product), then the "dot product and cross product combined"
is injective.
This combination, of course, is the tensor product. Specifically, when we're talking about something like $1/\vec{e_x}$, we want a thing whose tensor product with $\vec{e_x}$ has trace (dot product) 1 and commutator (wedge/cross product) 0, i.e. $\mathrm{tr}(\vec{e_x}'\vec{e_x})=1$ and $(\vec{e_x}'\vec{e_x})-(\vec{e_x}'\vec{e_x})^T=0$.
If all you've ever done in your life is Euclidean geometry, you'd probably think the answer to this question is $\vec{e_x}$ itself -- indeed, its dot product with $\vec{e_x}$ is 1 and its cross product with $\vec{e_x}$ is 0. But if you've ever done relativity and dealt with -- forget curved manifolds! -- the Minkowski manifold, you know that this is not necessarily true -- it depends on the metric tensor.
Could we
define a vector in a general co-ordinate system that is the inverse of $\vec{e_x}$? Yes, we can. But let's not do that (yet*) -- it just seems like there should be something more natural, or elegant, like we had with complex numbers.
So we define a space of "covectors", as "scalars divided by vectors" (informally speaking), call their basis $\tilde{e^i}$ which have the required dot and cross products. In Euclidean space -- and only in Euclidean space, these look exactly the same as vectors, and have exactly the same components. I like to call the conjugation here "metric conjugation", and the gradient is naturally a covector.
*As for the question of writing the gradient as a vector instead, this follows naturally using the metric tensor -- as an exercise, show, by considering the required vector corresponding to the covector $\tilde{e^x}$ (i.e. that has the right dot and cross products with $\vec{e_x}$) that the vector gradient can be given as the product of the inverse metric tensor and the covector gradient:
$${\partial ^\mu }f = {g^{\mu \nu }}{\partial _\nu }f$$
(Do this exercise! It is
the motivation for the metric tensor, and why it determines your co-ordinate system!)
I've been talking about the covector $\tilde{e^x}$ as being equal to the quotient "$1/\vec{e_x}$" but as I mentioned, this isn't really accurate -- the "1" in the quotient is a (1,1) tensor with trace 1 and commutator 0. Think about this tensor. Can you find this tensor in Clifford algebra? Maybe not. Can you find it as a linear transformation? Yes? Find it. And can you think of the covector alternatively as a quotient of a bivector and a trivector? Will you get $(e_y\wedge e_z)/(e_x\wedge e_y\wedge e_z)$?
