Every spot on a map is pinned down by exactly two numbers — how far east and how far north — and no fewer will ever do the job. That "exactly two" is no accident: it is the dimension of the plane, and the pair of directions those numbers are measured along is a basis. This page is about what those two words really mean.

Here's a question worth sitting with: what is the smallest set of ingredient vectors you'd ever need, so that every point in a space can be reached by some linear combination of them? Not "a set that works," but the leanest one possible — no wasted ingredients, nothing left unreachable. That minimal, non-redundant set is a basis, and how many vectors it takes is the space's dimension. The "2" in "2D" isn't just a label somebody picked — it's literally the size of a basis for the plane.

A set of vectors earns the title "basis" by doing exactly two jobs:

• spanning — their linear combinations reach every single point in the space, with nothing left out, and
• linearly independent — no vector in the set is redundant; none of them could be built from the others (the very trap you met when two vectors turned out to be parallel and collapsed onto a single line).

A basis is a perfect set of building blocks: just enough vectors to describe everything, with nothing redundant. The standard basis for 2D is \{\hat{\imath}, \hat{\jmath}\}, but as you'll see below, it is far from the only one — any two independent vectors will do. Step up to three dimensions and the same idea holds, just with one more ingredient: the standard basis becomes \{\hat{\imath}, \hat{\jmath}, \hat{k}\}, three vectors, because 3D space has dimension three.

Why insist on the smallest set? Because every extra, redundant vector you drag along has to be carried, stored, and computed with forever after — for no benefit, since it adds no new reach. A basis is the point where you've squeezed out every bit of redundancy but haven't lost any reach either: the leanest possible description of the whole space.

Worked examples: is it a basis?

1. The standard basis. Take \hat{\imath}=(1,0) and \hat{\jmath}=(0,1). Spanning: any point (x, y) is just x\,\hat{\imath} + y\,\hat{\jmath}, so every point is reachable. Independent: neither is a multiple of the other. Both boxes ticked — it's a basis.

2. A completely different basis. Try \vec{u}=(1,1) and \vec{v}=(1,-1) instead. Can they still reach every point? Take a target like (4, 2) and solve a\vec{u} + b\vec{v} = (4,2):

Check: 3(1,1) + 1(1,-1) = (4, 2) ✓. That trick — average the coordinates to get a, halve their difference to get b — works for any target point, so \{\vec{u}, \vec{v}\} spans the whole plane too. They aren't parallel, so they're independent. It's a different basis for the exact same 2D space — bases are not unique.

3. A set that fails. Now try \vec{u}=(2,1) and \vec{v}=(4,2). Notice \vec{v} = 2\vec{u} — they're parallel, so they're dependent, and every combination a\vec{u}+b\vec{v} simplifies to (2(a+2b),\, a+2b), which always lands on the line y = x/2. A point like (1, 0) is off that line, so it's simply unreachable. This pair spans only a line, not the plane — it fails the spanning test, so it's not a basis for 2D at all.

Quick self-check. Before reading on, decide for yourself: is \{(2,0),\,(0,3)\} a basis for the plane? It is — they're not parallel (independent), and any (x,y) is \tfrac{x}{2}(2,0) + \tfrac{y}{3}(0,3), so together they still reach every point (spanning), just with a stretched-out grid rather than a square one.

4. Too many vectors, guaranteed redundant. Take three 2D vectors at once: (1,0), (0,1), and (3,4). Even without checking anything fancy, we already know from the dimension of the plane that three vectors in 2D must be dependent — and sure enough,

so the third vector is just a combination of the first two: entirely redundant. Dropping it leaves \{(1,0),(0,1)\}, the standard basis, which alone already spans everything. This is exactly the "more vectors than the dimension are automatically dependent" rule in action.

A basis builds a grid

Each basis lays its own coordinate grid over the plane. Tilt the second basis vector below and watch the grid skew with it — yet as long as the two vectors stay independent, the grid still tiles the whole plane, so every point still has coordinates. Make them collinear and the grid collapses: no longer a basis.

Dimension: the magic number

Here is the remarkable fact: every basis for a given space has the same number of vectors. That count is the dimension of the space — 2 for the plane, 3 for the space we live in, and n for the lists of n numbers that machine learning runs on. Dimension is the honest measure of "how many independent directions are really here" — and it never depends on which basis you chose. Swap in the alternate basis \{(1,1),(1,-1)\} from the worked example above, and the plane is still exactly 2-dimensional — only the labelling of points changes, never the count.

Once you have a fixed basis, every point gets an honest address: its coordinates with respect to that basis are simply the weights you used to build it — the a=3, b=1 from worked example 2 above are the coordinates of (4,2) in the basis \{(1,1),(1,-1)\}. Different basis, different-looking coordinates, same underlying point.

Dimension isn't only for arrows in the plane, either. It's a completely general count of "independent knobs." The space of all 2\times 2 matrices, for instance, has dimension 4 — a basis is the four matrices with a single 1 in one slot and zeros elsewhere, since any 2\times 2 matrix is a unique combination of those four. The same counting idea works for polynomials, functions, and every other abstract vector space — wherever "linear combination" makes sense, "basis" and "dimension" come along for free.

Picking a well-suited basis is a running theme in linear algebra, not just a one-off trick. Later you'll meet eigenbases — special bases, tailored to a particular transformation, in which that transformation's action becomes as simple as stretching along each basis vector on its own. Same underlying idea as here: the right set of building blocks makes a hard problem easy.

See it explained