Dimension

Dimension is the single number that lets linear algebra talk about "size" at all, and every field that uses linear algebra leans on it. In numerical linear algebra, the dimension of a solution space (the kernel of a matrix) controls how many free parameters a linear system has, feeding directly into the rank–nullity theorem. In quantum mechanics, the state space of a physical system is a (finite- or infinite-dimensional) Vector Space over $\mathbb{C}$, and the dimension of the Hilbert space for a qubit ($2$), a qutrit ($3$), or a composite system of $n$ qubits ($2^n$) directly determines how much information the system can encode — the whole subject of quantum information theory tracks dimension counts obsessively. In coding theory and cryptography, a linear code is a subspace of $\mathbb{F}_q^n$, and its dimension $k$ (giving code parameters $[n,k]$) determines the information rate of the code; error-correcting capability and rate trade off precisely through dimension. In statistics, the dimension of the column space of a design matrix is the number of independent parameters a linear regression model can estimate, and degrees-of-freedom counts in ANOVA are dimension computations for nested subspaces. In computer graphics and robotics, the dimension of a configuration space (e.g., the space of joint angles of a robot arm, linearized locally) determines the degrees of freedom of a mechanism. In pure mathematics, dimension is the starting invariant for classifying finite-dimensional vector spaces up to isomorphism — two finite-dimensional spaces over the same field are isomorphic exactly when they have the same dimension, a fact that follows immediately once Coordinates Relative to a Basis are introduced.

:::quiz[Which hypothesis is essential in the Well-Definedness of Dimension theorem, and is violated by $\mathbb{R}[x]$ with basis $\{1,x,x^2,\dots\}$?]

• $V$ must be over an infinite field
• $V$ must have a finite basis ✓
• $V$ must be a subspace of $\mathbb{R}^n$
• The basis must be orthogonal

Ask the tutor: How does dimension combine with subspaces to give the dimension formulas $\dim(U+W) = \dim U + \dim W - \dim(U\cap W)$?

Think of a basis as a minimal, non-redundant "coordinate grid" for $V$: every vector in $V$ is a unique linear combination of the basis vectors, so once you fix a basis, every point of $V$ becomes a list of $n$ scalars. The number $n$ feels like it should be a property of the space, not of the grid — rotating the axes of $\mathbb{R}^3$ still gives three axes, never two or four. The exchange lemma is the precise mechanism that forbids a "four-axis grid" and a "two-axis grid" from coexisting: it lets you trade elements of one basis for elements of another, one at a time, without ever running out of room, which forces the two bases to match in size. Picture two bases of the same plane $V = \mathbb{R}^2$: $B_1 = \{(1,0),(0,1)\}$ and $B_2 = \{(1,1),(1,-1)\}$. Both have exactly two vectors. The exchange lemma says this is not a coincidence — it is a mathematical necessity, because $B_1$ is independent and $B_2$ spans (so $|B_1|\le|B_2|$), and symmetrically $B_2$ is independent and $B_1$ spans (so $|B_2|\le|B_1|$).

Example 1: Dimension of a subspace of $\mathbb{R}^4$ described by two independent constraints. Let

We compute $\dim W$ directly by exhibiting and verifying a basis, then invoking the theorem to know the answer is basis-independent. Solve the constraints. From the second equation, $x_4 = -x_2$. From the first, $x_1 = 2x_2 - x_3$. So every vector of $W$ has the form

with $x_2,x_3\in\mathbb{R}$ free. Let $v_1=(2,1,0,-1)$, $v_2=(-1,0,1,0)$. We have shown $\{v_1,v_2\}$ spans $W$. Independence: suppose $a v_1 + b v_2 = 0$. Comparing the second coordinate, $a\cdot 1 + b\cdot 0 = 0 \Rightarrow a=0$. Comparing the third coordinate, $a\cdot 0 + b\cdot 1 = 0 \Rightarrow b=0$. So $\{v_1,v_2\}$ is linearly independent. Thus $\{v_1,v_2\}$ is a basis of $W$, and by the well-definedness theorem this is the number of vectors in every basis of $W$:

As a check via Corollary (Independent sets, spanning sets, dimension): since $\dim W=2$, no independent subset of $W$ can have $3$ or more vectors. Indeed $v_3=(1,1,1,-1)=v_1+v_2\cdot(-1)$... let's verify directly: is $v_3 = (1,1,1,-1)$ a combination of $v_1,v_2$? We'd need $2a-b=1,\ a=1,\ b=1,\ -a=-1$. From $a=1,b=1$: check first equation $2(1)-1=1$ ✓, last $-1=-1$ ✓. So indeed $v_3=v_1+v_2$, confirming $v_3\in W$ but $\{v_1,v_2,v_3\}$ is automatically dependent, consistent with $\dim W = 2$.

Example 2: Dimension of the space of $2\times 3$ matrices with a trace-like linear constraint. Let $V = M_{2\times 3}(\mathbb{R})$, the space of all real $2\times 3$ matrices, and let

where $A = \begin{pmatrix} a_{11}&a_{12}&a_{13}\\ a_{21}&a_{22}&a_{23}\end{pmatrix}$. First, $\dim M_{2\times3}(\mathbb{R}) = 6$: the six matrix units $E_{ij}$ (having a $1$ in position $(i,j)$ and $0$ elsewhere, for $i\in\{1,2\}$, $j\in\{1,2,3\}$) are easily checked to be independent (a vanishing linear combination has each coefficient equal to the corresponding matrix entry, so all coefficients are $0$) and to span (every matrix is the combination of the $E_{ij}$ with coefficients equal to its own entries). Hence they form a basis, so $\dim M_{2\times 3}(\mathbb{R})=6$ by the theorem. Now for $U$: it is the solution set of one linear equation in the $6$ free entries $a_{11},a_{12},a_{13},a_{21},a_{22},a_{23}$, namely $a_{11}+2a_{12}-a_{23}=0$, i.e. $a_{11} = -2a_{12}+a_{23}$. Every matrix in $U$ is determined by choosing $a_{12},a_{13},a_{21},a_{22},a_{23}$ freely (5 free scalars) and setting $a_{11}=-2a_{12}+a_{23}$. So a natural spanning set for $U$ is obtained by setting each free variable to $1$ in turn and the rest to $0$:

(These come from setting $(a_{12},a_{13},a_{21},a_{22},a_{23})$ equal to the five standard basis vectors of $\mathbb{R}^5$ and reading off $a_{11}$ each time.) Spanning: every $A \in U$ equals $a_{12}B_1 + a_{13}B_2 + a_{21}B_3 + a_{22}B_4 + a_{23}B_5$ — check the $(1,1)$-entry of this combination is $-2a_{12}+a_{23}=a_{11}$ as required, and all other entries match by inspection. Independence: if $c_1B_1+\cdots+c_5B_5 = 0$, then comparing entries $(1,2),(1,3),(2,1),(2,2),(2,3)$ directly gives $c_1=0,c_2=0,c_3=0,c_4=0,c_5=0$ (each of these five positions is touched by only one $B_i$). So $\{B_1,\dots,B_5\}$ is a basis of $U$, and by the theorem,

This matches the general principle (a foreshadow of dimension formulas) that imposing one nontrivial linear constraint drops dimension by exactly $1$.