What Is $\partial_i$? Basis Vectors of the Tangent Space
1. Why Can a Derivative Serve as a Vector?
In Euclidean space, we are used to thinking of vectors as arrows. On a curved manifold, arrows cannot be drawn globally, so tangent vectors must be defined intrinsically. The entry point is the directional derivative.
Take a curve $\gamma(t)$ through a point $p$, with $\gamma(0)=p$ and coordinates $x^i(t)$. The tangent vector of the curve at $p$ should encode its direction and speed. Differentiate a function $f$ along the curve, using the summation convention from Foundations 2:
The right-hand side is a combination of partial derivatives of $f$, weighted by the coordinate velocities $dx^i/dt$. This suggests that the partial-derivative operators themselves can serve as basis vectors, with the coordinate velocities as components.
So we define $\partial_i|_p$ as the operator that takes a function $f$, written in terms of the coordinates $x^1,\dots,x^n$, and returns its ordinary partial derivative with respect to $x^i$ at $p$:
The $n$ operators $\partial_1|_p, \dots, \partial_n|_p$ form a basis of the $n$-dimensional tangent space $T_pM$. Any tangent vector can be written:
2. $\partial_x$ and $e_x$: Same Vector, Two Languages
You may have met the unit arrow $e_x$ (or $e_1$) in linear algebra or physics. Is $\partial_x$ the same thing?
In Cartesian coordinates on the plane, yes. Move along the straight path $\gamma(t) = p + t\,e_x$. Its velocity is the arrow $e_x$, and the rate of change of a function $f$ along it is:
So the arrow $e_x$ and the operator $\partial_x$ are the same vector, written in two languages. The arrow notation emphasizes length and direction. The operator notation emphasizes what the vector does to functions.
| Notation | Where it is common | Emphasizes |
|---|---|---|
| $e_x,\ e_1$ | Linear algebra, physics, Euclidean geometry | An arrow, with length and direction |
| $\partial_x,\ \partial_i$ | Differential geometry, general relativity | An operator, the directional derivative |
Only the second language still works when the arrow picture breaks down, as we see next.
3. Why Arrows Are Not Enough on a Curved Space
On a flat plane, an arrow can be drawn on the plane itself. The plane is its own tangent space, so arrows at different points can be moved around and compared freely.
On a sphere, such as the surface of the Earth, it is different. Take the arrow that points north at San Francisco. As a straight arrow, it lies in the tangent plane, which touches the sphere at a single point and sticks out into the surrounding three-dimensional space. The arrow is not on the surface. (An arrow drawn along the surface would follow a curve, not a straight line.) So describing a tangent vector as an arrow borrows the space outside.
But an ant living on the surface cannot see the outside. It still knows what it means to walk north: it can follow a path through its point and ask how quickly things change along it. That is the intrinsic definition. A tangent vector is the velocity of a path, or equivalently the directional-derivative operator that the path defines:
Nothing here refers to an outside space. This matters for two reasons. Curvature is intrinsic, as Gauss showed, so it must be describable from within. And spacetime in general relativity has no outside to borrow from.
So in flat space, $\partial_i$ is not a different vector from $e_i$. It is the same vector, described in a way that survives when the arrow picture breaks down.
4. The Coordinate Curve Picture
For coordinates $x^1,\dots,x^n$, the $i$-th coordinate curve is the curve along which only $x^i$ changes and all the others stay fixed. Its tangent direction is $\partial_i$. Intuitively, $\partial_i$ points in the direction in which only the $i$-th coordinate changes.
Stand at $p$ and move along the $x^1$ direction, and $f$ changes at the rate $\partial_1 f$. Move along $x^2$, and the rate is $\partial_2 f$. These $n$ directional derivatives are independent, since applying $\partial_i$ to a coordinate function gives
and they span the tangent space, because any tangent direction can be broken into a combination of the coordinate directions.
5. Behavior Under Coordinate Changes
This is the link back to Foundations 1. Change to new coordinates $\tilde{x}^i$. The chain rule gives, for any function $f$:
Since $f$ is arbitrary, the operators themselves obey:
The basis vectors transform by the matrix $A = \partial x/\partial\tilde{x}$, while the components transform by its inverse, $\tilde{v}^i = \frac{\partial \tilde{x}^i}{\partial x^j}v^j$. That is exactly the opposite behavior from Foundations 1. In the language of Foundations 2, the basis vectors carry a lower index and transform like covariant components, while the vector components carry an upper index and are contravariant. The combination stays fixed:
6. A Common Confusion: $\partial_i$ Is Not a Small Displacement
It is tempting to picture $\partial_i$ as a tiny displacement of length one along the $x^i$ direction. That picture works in Cartesian coordinates, where the coordinate basis is orthonormal. It fails in curvilinear coordinates, even in a flat plane. The length and angles of the basis vectors are set by the metric:
In polar coordinates on the plane, the metric $ds^2 = dr^2 + r^2d\theta^2$ gives:
So $\partial_r$ has length $1$, but $\partial_\theta$ has length $r$. It is a tangent vector, a velocity, not an arc-length element. So $\partial_i$ is best thought of as an operator, the directional derivative along the coordinate curve. Only in Cartesian coordinates on flat space do the operator and the unit arrow coincide.
The physicist’s unit vector $e_\theta$ is the normalized version of $\partial_\theta$. It points the same way but has length $1$:
The two agree only when the coordinate basis is already orthonormal, as in Cartesian coordinates.
7. A Subscript You Should Not Ignore
In $\partial_i|_p$, the point $p$ is easy to overlook, but it matters. The basis at $p$ lives in the tangent space $T_pM$, and the basis at another point $q$ lives in $T_qM$. In polar coordinates, for example, $\partial_\theta$ points in a different direction at every point and gets longer as $r$ increases. These are different spaces, so we cannot simply add or subtract vectors from them.
How the basis changes from point to point is the subject of Foundations 5, on Christoffel symbols.
8. Where Does This Lead?
Once we accept that tangent vectors are directional-derivative operators, the next natural question is: what is the object that eats a vector and returns a number? That is the covector, and its coordinate basis is $dx^i$. We glimpsed the key relationship in Foundations 2:
This is the dual basis relationship. The symbol $dx^i$ is not an infinitesimal quantity. It is a linear functional that extracts the $i$-th component of a vector.