Linear Independence, Basis, and Dimension

Linear algebra earns its power the moment a vector space stops being an abstract fog of addable, scalable objects and acquires a skeleton — a short, honest list of vectors from which every other vector is built in exactly one way. That list is a basis, and choosing one is the difference between a space you can only describe and a space you can compute in. Two demands pin it down. The list must be economical: no member may be a disguised combination of the others, or it is dead weight. And it must be complete: nothing in the space may lie beyond its reach. Economy is the content of linear independence, completeness is spanning, and a basis is the rare list that is both at once — a minimal spanning set and, equivalently, a maximal independent set. The theorem that makes the whole subject rigid is that every basis of a given space has the same number of vectors. That number, the dimension, is an intrinsic invariant: it does not know or care which basis you happened to pick. Dimension counts genuine degrees of freedom — $\mathbb{R}^3$ has dimension three because naming a point needs three numbers, never two and never four. And once a basis is fixed, every vector acquires a unique tuple of coordinates, the bridge that turns coordinate-free arguments into columns of numbers and turns linear operators into matrices. From here the rest of linear algebra unfolds: rank and nullity, change of basis and similarity, eigen-decompositions, and the orthonormal bases of quantum mechanics. This article builds that skeleton and proves it holds together.

Key takeaways

A set of vectors is linearly independent when the only combination equal to the zero vector is the all-zero one; equivalently, no vector is a combination of the others.
A set spans a space when every vector is some combination of it; a basis is a set that is simultaneously independent and spanning.
Every basis of a fixed space has the same number of elements — that invariant count is the dimension of the space.
Relative to a chosen basis each vector has unique coordinates, which is exactly what licenses replacing abstract vectors by column vectors and operators by matrices.
The Grassmann formula $\dim(U+W)=\dim U+\dim W-\dim(U\cap W)$ governs how subspaces combine, and a nonzero determinant certifies a basis of $\mathbb{F}^n$ .
Key idea: A basis is at once a minimal spanning set and a maximal linearly independent set; its size — the dimension — is the same for every basis and equals the number of independent coordinates needed to name a vector.

The core insight

A basis is simultaneously a minimal spanning set and a maximal linearly independent set; its size, the dimension, is the same for every basis and equals the number of independent coordinates needed to name a vector.

Where this lives today

Dimension-counting is not a museum piece — it is the accounting that keeps modern physics and computation honest. A register of $n$ qubits lives in a Hilbert space of dimension $2^n$ , and that single exponential is the whole promise and difficulty of quantum computing: fifty qubits already exceed a petabyte of amplitudes, so the state cannot be written down classically. Quantum error correction is pure subspace geometry — a stabilizer code hides logical information in a carefully chosen subspace, and an $[[n,k,d]]$ code encodes $k$ logical qubits into $n$ physical ones by reserving a $2^k$ -dimensional code subspace inside a $2^n$ -dimensional register. Classical linear codes are the same idea over $\mathbb{F}_2$ : an $[n,k]$ code is a $k$ -dimensional subspace, and its dimension is literally the number of message bits it carries. In data science, principal-component analysis and every low-rank method estimate the effective dimension of a cloud of measurements, discarding near-dependent directions to compress and denoise. The finite-element and spectral methods that simulate everything from bridges to plasmas work by choosing a finite-dimensional basis of trial functions and solving in coordinates. And in physics proper, the count of normal modes, of independent tensor components allowed by a symmetry, or of degrees of freedom in a gauge theory is always, underneath, a dimension of a vector space — Grassmann's invariant doing quiet bookkeeping.

The physical picture

A basis is best pictured as a complete, non-redundant set of directions. Complete means every point in the space is reachable by mixing those directions; non-redundant means deleting any one of them shrinks what you can reach. Crucially, the directions need not be perpendicular and need not be unit length — a skew, stretched pair of arrows in the plane is every bit as good a basis as the tidy grid, because "basis" is about independence and reach, not about angles or lengths. The first figure makes the coordinate idea concrete: a slanted basis still assigns every vector one and only one coordinate pair.

Figure

A skew basis of ℝ²: the independent vectors e₁ and e₂ generate the whole plane, and every point gets a unique coordinate pair — here v = 3e₁ + 2e₂, so v carries the coordinates (3, 2) even though the basis is neither orthogonal nor normalized

Linear independence, by contrast, is a statement about a single vector — the zero vector. The vectors are independent precisely when the only recipe that returns you to the origin is to take none of each. If some clever nonzero recipe also lands on the origin, then one vector is secretly a combination of the rest and can be thrown away without loss. This is why counting is so rigid: you can never cram more than $\dim V$ independent vectors into a space, nor span it with fewer. That rigidity is exactly what lets dimensions of subspaces be added and subtracted like sizes of sets, corrected for their overlap — the Grassmann formula, illustrated below with two planes in ordinary three-space.

Figure

Two distinct planes through the origin in ℝ³ each have dimension 2 and meet in a line of dimension 1, so the Grassmann formula gives dim(U + W) = 2 + 2 − 1 = 3 and their sum fills all of ℝ³

The core equations

Everything begins with a test on the zero vector. A finite list of vectors is linearly independent when the vanishing of a linear combination forces every coefficient to vanish — there is no nontrivial way to cancel the vectors against one another. This single implication is the whole definition, and every later structural fact is squeezed out of it.

\sum_{i=1}^{n} c_i \mathbf{v}_i = \mathbf{0} \;\Longrightarrow\; c_1 = c_2 = \cdots = c_n = 0 \qquad (1)

where: $\mathbf{v}_i$ — the vectors under test, living in a space $V$ over a field $\mathbb{F}$ ; $c_i$ — scalars drawn from $\mathbb{F}$ ; $\mathbf{0}$ — the zero vector of $V$ . Read it as: the origin can be written in only the trivial way, so no vector is redundant. If instead some nonzero coefficient choice hits $\mathbf{0}$ , the set is dependent and one member is expendable. Spanning is the complementary reach condition: the span of a set is the collection of all its finite linear combinations, and the set spans $V$ when that collection is the whole of $V$ .

\operatorname{span}\{\mathbf{v}_1,\dots,\mathbf{v}_n\} = \Big\{ \textstyle\sum_{i=1}^{n} c_i \mathbf{v}_i : c_i \in \mathbb{F} \Big\} \qquad (2)

where: $\operatorname{span}$ — the smallest subspace containing the listed vectors; $c_i$ — arbitrary scalars. Read it as: nothing in $V$ is out of reach. Independence and spanning are logically separate — a set can span while being bloated with redundancy, or be lean and independent yet fail to reach everything. A basis is exactly the list that passes both tests at once, so it is lean and complete together.

\mathcal{B} = \{\mathbf{e}_1,\dots,\mathbf{e}_n\} \text{ is a basis} \iff \mathcal{B}\text{ is independent and spans } V \qquad (3)

where: $\mathcal{B}$ — an ordered list of vectors; $n$ — its size, which will turn out to be $\dim V$ . Read it as: a basis is a minimal spanning set (drop one and spanning fails) and, equivalently, a maximal independent set (add one and independence fails). Because a basis both spans and is independent, every vector expands over it in one and only one way, and the coefficients of that expansion are the vector's coordinates.

\mathbf{v} = \sum_{i=1}^{n} v_i\, \mathbf{e}_i, \qquad [\mathbf{v}]_{\mathcal{B}} = (v_1,\dots,v_n)^{\mathsf{T}} \qquad (4)

where: $v_i$ — the $i$ -th coordinate of $\mathbf{v}$ in $\mathcal{B}$ ; $[\mathbf{v}]_{\mathcal{B}}$ — the coordinate column vector. Read it as: fixing a basis installs a faithful, invertible dictionary $\mathbf{v}\leftrightarrow[\mathbf{v}]_{\mathcal{B}}$ between abstract vectors and columns of numbers — the coordinate isomorphism $V\cong\mathbb{F}^n$ that Derivation B secures. The dimension is the size shared by every basis. That there is such a shared size — that no two bases can disagree — is the load-bearing theorem, proved in Derivation A.

\dim V = |\mathcal{B}| = n \qquad (\text{same integer for every basis}) \qquad (5)

where: $|\mathcal{B}|$ — the number of vectors in a basis; $n$ — the dimension, an intrinsic invariant of $V$ . Read it as: dimension is a property of the space, not of any chart on it. A subspace $U\subseteq V$ obeys $\dim U\le\dim V$ , with equality only when $U=V$ — you cannot fit a bigger independent set inside a smaller room. When two subspaces are combined, their dimensions add, but the overlap has been counted twice and must be subtracted once — the inclusion–exclusion of linear algebra.

\dim(U+W) = \dim U + \dim W - \dim(U\cap W) \qquad (6)

where: $U+W=\{\mathbf{u}+\mathbf{w}:\mathbf{u}\in U,\ \mathbf{w}\in W\}$ — the sum subspace; $U\cap W$ — the intersection subspace. Read it as: overlap is charged exactly once. When $U\cap W=\{\mathbf{0}\}$ the sum is direct, $\dim(U\oplus W)=\dim U+\dim W$ , and every element splits uniquely into a $U$ -part and a $W$ -part. Finally, in the concrete space $\mathbb{F}^n$ there is a one-line computational test: stack $n$ candidate vectors as the columns of a square matrix and take its determinant.

\det[\,\mathbf{v}_1\;\cdots\;\mathbf{v}_n\,] \ne 0 \iff \{\mathbf{v}_1,\dots,\mathbf{v}_n\} \text{ is a basis of } \mathbb{F}^n \qquad (7)

where: $\det$ — the determinant of the matrix whose columns are the $\mathbf{v}_i$ . Read it as: a nonzero determinant certifies invertibility, hence independence; and in $\mathbb{F}^n$ any $n$ independent vectors automatically span, so they form a basis. A zero determinant flags a hidden dependence among the columns.

Derivation AInvariance of dimension by the Steinitz exchange

The goal is to prove the single fact on which the entire notion of dimension rests: that no linearly independent set can be larger than any spanning set of the same space. Once that inequality is in hand, applying it twice — with the roles of two bases swapped — forces any two bases to have exactly the same number of vectors, so the integer we call the dimension is unambiguous. The engine is a replacement procedure: feed independent vectors into a spanning set one at a time, each time evicting a spanning vector to keep the count fixed, and show the eviction never stalls.

Step 1. Let $L=\{\mathbf{u}_1,\dots,\mathbf{u}_m\}$ be independent and $S=\{\mathbf{w}_1,\dots,\mathbf{w}_n\}$ span $V$ ; we aim to show $m\le n$ . Since $S$ spans, write $\mathbf{u}_1=\sum_j a_j\mathbf{w}_j$ , and because $\mathbf{u}_1\ne\mathbf{0}$ some coefficient $a_k\ne 0$ . Why: an independent set contains no zero vector, so at least one spanning vector genuinely participates in building $\mathbf{u}_1$ , giving us something to solve for.

Step 2. Solve that relation for $\mathbf{w}_k=a_k^{-1}\!\big(\mathbf{u}_1-\sum_{j\ne k}a_j\mathbf{w}_j\big)$ and delete $\mathbf{w}_k$ from the list, inserting $\mathbf{u}_1$ in its place. The new set $\{\mathbf{u}_1\}\cup(S\setminus\{\mathbf{w}_k\})$ still spans $V$ and still has exactly $n$ elements. Why: anything $\mathbf{w}_k$ used to build is now rebuildable from $\mathbf{u}_1$ and the survivors, so reach is preserved while the head-count is held fixed — this is the "exchange".

Step 3. Repeat with $\mathbf{u}_2$ : express it in the current spanning set. The coefficients of the remaining $\mathbf{w}$ 's cannot all be zero, for otherwise $\mathbf{u}_2$ would be a combination of $\mathbf{u}_1$ alone, contradicting the independence of $L$ . So some $\mathbf{w}$ can again be evicted in favour of $\mathbf{u}_2$ . Why: independence is precisely what guarantees a fresh spanning vector is always available to swap out — the process can never be forced to consume one of the already-placed $\mathbf{u}$ 's.

Step 4. After $t$ exchanges, $\{\mathbf{u}_1,\dots,\mathbf{u}_t\}$ together with $n-t$ of the original $\mathbf{w}$ 's spans $V$ . The procedure can only continue while spanning vectors remain to be evicted, so the number of $\mathbf{u}$ 's cannot exceed the number of $\mathbf{w}$ 's: $m\le n$ . Why: each inserted independent vector costs one spanning vector, and you cannot spend more than you started with — that is the whole counting bound.

Step 5. Now let $\mathcal{B}_1,\mathcal{B}_2$ both be bases. Each is independent and each spans, so applying the bound both ways gives $|\mathcal{B}_1|\le|\mathcal{B}_2|$ and $|\mathcal{B}_2|\le|\mathcal{B}_1|$ , hence $|\mathcal{B}_1|=|\mathcal{B}_2|$ . Why: the common value is basis-independent by construction, which is exactly what entitles us to name it $\dim V$ .

Go deeper · Derivation B · 1 minWhy a basis gives every vector unique coordinates

Derivation BWhy a basis gives every vector unique coordinates

The aim is to show that a basis does more than reach every vector — it reaches each one along a single, unrepeatable route, so the map that sends a vector to its tuple of coefficients is well defined, one-to-one, and linear. This is the theorem that turns geometry into arithmetic: it certifies that replacing $\mathbf{v}$ by its coordinate column $[\mathbf{v}]_{\mathcal{B}}$ loses no information and respects all vector-space operations, so any question about $V$ can be pushed to $\mathbb{F}^n$ and answered with matrices. Existence of coordinates is spanning; uniqueness is independence; together they give an isomorphism.

Step 1. Existence: because $\mathcal{B}=\{\mathbf{e}_1,\dots,\mathbf{e}_n\}$ spans, every $\mathbf{v}\in V$ can be written as $\mathbf{v}=\sum_i v_i\mathbf{e}_i$ for some scalars $v_i$ . Why: spanning is exactly the guarantee that a coordinate tuple exists at all — without it some vectors would have no address in the basis.

Step 2. Uniqueness: suppose $\mathbf{v}=\sum_i v_i\mathbf{e}_i=\sum_i v_i'\mathbf{e}_i$ are two expansions. Subtracting gives $\sum_i (v_i-v_i')\mathbf{e}_i=\mathbf{0}$ , and independence forces every coefficient to vanish, so $v_i=v_i'$ for all $i$ . Why: two different addresses for one vector would be a nontrivial combination summing to zero — precisely what independence outlaws.

Step 3. Define $\Phi:V\to\mathbb{F}^n$ by $\Phi(\mathbf{v})=[\mathbf{v}]_{\mathcal{B}}=(v_1,\dots,v_n)^{\mathsf{T}}$ . Steps 1–2 make $\Phi$ a well-defined bijection, and expanding $\alpha\mathbf{v}+\beta\mathbf{w}$ over $\mathcal{B}$ shows $\Phi(\alpha\mathbf{v}+\beta\mathbf{w})=\alpha\Phi(\mathbf{v})+\beta\Phi(\mathbf{w})$ . Why: a bijection that respects addition and scaling is a linear isomorphism $V\cong\mathbb{F}^n$ , so every $n$ -dimensional space over $\mathbb{F}$ is a faithful copy of coordinate space — the formal license for all matrix computation.

Go deeper · Derivation C · 2 minThe Grassmann dimension formula for a sum of subspaces

Derivation CThe Grassmann dimension formula for a sum of subspaces

The target is the inclusion–exclusion law $\dim(U+W)=\dim U+\dim W-\dim(U\cap W)$ , which tells you how much room two subspaces occupy together once their shared part is discounted. The strategy is constructive and is a template worth remembering: start from a basis of the overlap, extend it separately into each subspace, glue the three pieces, and prove the glued set is a basis of the sum by checking it both spans and is independent. Counting its members then reads off the formula directly. Let $\dim(U\cap W)=k$ , $\dim U=p$ , $\dim W=q$ .

Step 1. Choose a basis $\{\mathbf{x}_1,\dots,\mathbf{x}_k\}$ of the intersection $U\cap W$ . Because these vectors lie in both $U$ and $W$ , they can be extended in either direction. Why: anchoring on the shared part is what prevents double-counting — the overlap gets one basis, used by both sides.

Step 2. Extend to a basis $\{\mathbf{x}_i\}\cup\{\mathbf{u}_1,\dots,\mathbf{u}_{p-k}\}$ of $U$ and, separately, to a basis $\{\mathbf{x}_i\}\cup\{\mathbf{w}_1,\dots,\mathbf{w}_{q-k}\}$ of $W$ , adding $p-k$ and $q-k$ fresh vectors respectively. Why: Derivation A's extension procedure guarantees these completions exist, and the added vectors carry the parts of $U$ and $W$ that stick out beyond the overlap.

Step 3. Form the pooled set $\{\mathbf{x}_i\}\cup\{\mathbf{u}_a\}\cup\{\mathbf{w}_b\}$ . It plainly spans $U+W$ , since any $\mathbf{u}+\mathbf{w}$ expands over the two extended bases, and it has $k+(p-k)+(q-k)=p+q-k$ members. Why: spanning is immediate; the entire content of the theorem is whether these $p+q-k$ vectors are independent, with no hidden redundancy across the seam.

Step 4. Test independence: suppose $\sum_i\alpha_i\mathbf{x}_i+\sum_a\beta_a\mathbf{u}_a+\sum_b\gamma_b\mathbf{w}_b=\mathbf{0}$ . Then $\sum_b\gamma_b\mathbf{w}_b=-\sum_i\alpha_i\mathbf{x}_i-\sum_a\beta_a\mathbf{u}_a$ lies in $U$ ; but its left side lies in $W$ , so it lies in $U\cap W$ and is a combination of the $\mathbf{x}_i$ alone. Independence within the $W$ -basis then forces every $\gamma_b=0$ , after which the leftover relation forces every $\alpha_i=\beta_a=0$ . Why: the vector caught in the intersection has two incompatible descriptions unless all the cross-terms vanish — that squeeze is exactly where the $-\dim(U\cap W)$ correction is born.

Step 5. The pooled set is therefore a basis of $U+W$ , and its size is the dimension: $\dim(U+W)=p+q-k$ . Why: with spanning from Step 3 and independence from Step 4, counting members gives the formula with no further work.

Limiting cases

The abstract theorems collapse to familiar tallies in the spaces physicists meet daily; each row is just $\dim V$ read off from the obvious basis.

Case	Result	Meaning
Coordinate space $\mathbb{F}^n$	$\dim\mathbb{F}^n=n$	The standard basis $\mathbf{e}_i$ with $(\mathbf{e}_i)_j=\delta_{ij}$ is the prototype; every column vector is its own coordinate tuple.
Polynomials of degree $\le n$	$\dim\mathcal{P}_n=n+1$	The monomials $\{1,x,x^2,\dots,x^n\}$ are independent and span — note it is $n+1$ , not $n$ , because the constant counts.
$m\times n$ matrices	$\dim\mathbb{F}^{m\times n}=mn$	The matrix units $E_{ij}$ (a single $1$ at position $i,j$ ) form the basis, so operators on $\mathbb{F}^n$ live in an $n^2$ -dimensional space.
Solutions of $y''+\omega^2 y=0$	$\dim=2$	A second-order linear ODE has a two-dimensional solution space, basis $\{\cos\omega t,\sin\omega t\}$ — two constants of integration, two dimensions.
Infinite-dimensional spaces	$\dim V=\infty$	All polynomials, or $\ell^2$ : no finite set spans. Hamel bases exist but are non-constructive, which is why Hilbert spaces use orthonormal topological bases instead.

Common trapA set that spans must be independent, and an independent set must span

Neither implication holds — spanning and independence are genuinely separate conditions, and conflating them is the single most common error in the subject. A set can span $\mathbb{R}^2$ while being dependent (three vectors, say: they reach everything but carry a redundancy), and a set can be independent while failing to span (a single nonzero vector in $\mathbb{R}^2$ : no redundancy, but it only reaches a line). Only when both hold at once do you have a basis, and only then is the count guaranteed to equal $\dim V$ . Two guardrails follow immediately. First, the zero vector can never belong to an independent set, because $1\cdot\mathbf{0}=\mathbf{0}$ is already a nontrivial dependence. Second, any list of more than $\dim V$ vectors is automatically dependent, and any list of fewer than $\dim V$ automatically fails to span — so the only candidate bases have exactly $\dim V$ members, and for a set of that exact size, independence and spanning become equivalent (checking either one suffices).

Worked examples

1 · Certifying a basis of ℝ³ with a determinant (fully worked)

Do the vectors $\mathbf{v}_1=(1,2,3)$ , $\mathbf{v}_2=(0,1,4)$ , $\mathbf{v}_3=(5,6,0)$ form a basis of $\mathbb{R}^3$ ?

Step 1. Three vectors in a three-dimensional space are a basis exactly when they are independent, and equation (7) turns that into a determinant. Stack them as the columns of a matrix:

A=\begin{pmatrix} 1 & 0 & 5\\ 2 & 1 & 6\\ 3 & 4 & 0 \end{pmatrix}.

Step 2. Expand along the first row: $\det A = 1\,(1\cdot 0-6\cdot 4) - 0\,(2\cdot 0-6\cdot 3) + 5\,(2\cdot 4-1\cdot 3) = 1(-24) - 0 + 5(5) = -24+25 = \mathbf{1}$ .

Step 3. Since $\det A=1\ne 0$ , the matrix is invertible, so its columns are linearly independent. In $\mathbb{R}^3$ three independent vectors automatically span, hence the set is a basis — and because $\det A=1$ , the parallelepiped they span has unit volume, the same as the standard basis. Had the determinant come out zero, the three vectors would have been coplanar (dependent) and would have spanned only a plane. Result: det = 1 ≠ 0, so {v₁, v₂, v₃} is a basis of ℝ³.

2 · Coordinates in a non-standard basis (fully worked)

The set $\mathcal{B}=\{\mathbf{b}_1,\mathbf{b}_2\}$ with $\mathbf{b}_1=(1,1)$ and $\mathbf{b}_2=(1,-1)$ is a basis of $\mathbb{R}^2$ . Find the coordinates of $\mathbf{v}=(3,1)$ in $\mathcal{B}$ .

Step 1. We seek scalars $a,b$ with $a\mathbf{b}_1+b\mathbf{b}_2=\mathbf{v}$ , i.e. $a(1,1)+b(1,-1)=(3,1)$ . Matching components gives the system $a+b=3$ and $a-b=1$ .

Step 2. Add the equations: $2a=4\Rightarrow a=2$ . Subtract them: $2b=2\Rightarrow b=1$ . So $[\mathbf{v}]_{\mathcal{B}}=(2,1)^{\mathsf{T}}$ . Check: $2(1,1)+1(1,-1)=(3,1)=\mathbf{v}$ . ✓

Step 3. This basis happens to be orthogonal, since $\mathbf{b}_1\cdot\mathbf{b}_2=1-1=0$ , so the same coordinates drop out of projections without solving any system: $a=\dfrac{\mathbf{v}\cdot\mathbf{b}_1}{\mathbf{b}_1\cdot\mathbf{b}_1}=\dfrac{4}{2}=2$ and $b=\dfrac{\mathbf{v}\cdot\mathbf{b}_2}{\mathbf{b}_2\cdot\mathbf{b}_2}=\dfrac{2}{2}=1$ . The two routes agree, previewing why orthonormal bases (Gram–Schmidt) are so convenient. Result: [v]_B = (2, 1); the same vector (3, 1) that reads (3, 1) in the standard basis reads (2, 1) here.

3 · The Grassmann formula on two planes

In $\mathbb{R}^3$ let $U$ be the $xy$ -plane, $U=\operatorname{span}\{(1,0,0),(0,1,0)\}$ , and $W$ the $yz$ -plane, $W=\operatorname{span}\{(0,1,0),(0,0,1)\}$ . Find $\dim(U+W)$ .

Step 1. Each plane is two-dimensional: $\dim U=\dim W=2$ . Their intersection consists of vectors in both planes, which must have zero $z$ -component (to sit in $U$ ) and zero $x$ -component (to sit in $W$ ), leaving the $y$ -axis: $U\cap W=\operatorname{span}\{(0,1,0)\}$ , so $\dim(U\cap W)=1$ .

Step 2. Apply equation (6): $\dim(U+W)=\dim U+\dim W-\dim(U\cap W)=2+2-1=\mathbf{3}$ .

Step 3. A three-dimensional subspace of $\mathbb{R}^3$ must be all of $\mathbb{R}^3$ , so $U+W=\mathbb{R}^3$ — the two planes together reach every point, as the second figure shows. Naively adding $2+2=4$ would exceed $\dim\mathbb{R}^3=3$ ; the shared $y$ -axis, counted once in each plane, is the double-count the formula removes. Result: dim(U + W) = 2 + 2 − 1 = 3, i.e. U + W = ℝ³.

Physics of the portrait

The ideas of independence, span, and dimension look timeless, but they were invented — assembled almost single-handedly, and decades ahead of their audience, by a Prussian schoolteacher who could not get mathematicians to read him.

Origins · 1 minHermann Grassmann and how this was discovered

Hermann Grassmann

1809–1877 · the Ausdehnungslehre and the birth of linear algebra

Hermann Günther Grassmann taught mathematics and physics at a Gymnasium in Stettin and never held a university chair, yet in his 1844 Die lineale Ausdehnungslehre ("The Theory of Linear Extension") he built, from scratch, essentially the entire apparatus of this article: linearly independent quantities, the span of a system, subspaces, the dimension of a space, and even the sum-and-intersection formula that now bears his name. He worked in arbitrary dimension at a time when mathematics rarely ventured past three, and he did it in a dense, philosophical prose that repelled contemporaries — Möbius declined to review it, and the book sold so poorly the publisher pulped the unsold stock. Discouraged, Grassmann turned to linguistics, where his work on Sanskrit and his phonetic law of Indo-European aspirates ("Grassmann's law") made him famous first. A drastically rewritten 1862 edition fared little better in his lifetime. Only near his death did figures like Hankel, Clebsch, and later Peano recognize what he had done; Peano's crisp 1888 axioms for a "linear system" are a direct descendant, and Gibbs and others mined the Ausdehnungslehre for vector and exterior algebra. The modern verdict is unambiguous: Grassmann is the founder of linear algebra as a coordinate-free theory, and the dimension formula, the exterior product, and the very notion of an $n$ -dimensional vector space are his intellectual children.

History

1844

— Grassmann's Ausdehnungslehre introduces independence, span, subspaces, and dimension

1862

— Grassmann issues a fully rewritten second edition, still largely unread in his lifetime

1888

— Peano gives the modern axioms of a vector space in Calcolo geometrico

1913

— Steinitz sharpens the exchange lemma that pins down the invariance of dimension

1918

— Weyl's Raum, Zeit, Materie carries the axiomatic vector-space viewpoint into physics

Going further

Texts. Axler, Linear Algebra Done Right, 4th ed., Ch. 2 (spanning, independence, bases, dimension) · Halmos, Finite-Dimensional Vector Spaces, §§5–9 · Stone & Goldbart, Mathematics for Physics, Ch. 1 · Arfken & Weber, Mathematical Methods for Physicists, 7th ed., Ch. 3 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 8 · Hassani, Mathematical Physics, Ch. 2. The key numerical results in this article are: the standard dimensions $\dim\mathbb{F}^n=n$ , $\dim\mathcal{P}_n=n+1$ , $\dim\mathbb{F}^{m\times n}=mn$ ; the determinant $\det A=1$ of Example 1; the coordinates $[\mathbf{v}]_{\mathcal{B}}=(2,1)$ of Example 2 (obtainable by both elimination and projection); and the Grassmann count $2+2-1=3$ of Example 3. No empirical constants appear — every number here is a theorem, not a measurement.

Prerequisites. Vector Spaces and Subspaces. Leads to. Linear Operators and Matrix Representations · Change of Basis and Similarity Transformations · Rank, Nullity, and the Fundamental Theorem of Linear Maps · Orthogonality and the Gram–Schmidt Process.

Ask the tutor: Ask me to row-reduce a specific set of vectors to read off its rank and dependence relations, or to show how the coordinate isomorphism $V\cong\mathbb{F}^n$ turns a linear operator into its matrix.

What is the difference between a basis and a spanning set?

A spanning set only guarantees reach — every vector is some combination of it — but it may be padded with redundant vectors, so a vector can have many different expansions. A basis is a spanning set with the padding removed: it is spanning and independent, which upgrades "some expansion exists" to "exactly one expansion exists". Practically, you get a basis from a spanning set by pruning any vector that is a combination of the others, repeating until nothing is redundant. Every basis is a spanning set, but only the minimal spanning sets are bases.

Can an infinite-dimensional space have a basis?

It depends on which notion you mean, and this is a genuine subtlety physicists must respect. In the purely algebraic sense, every vector space has a Hamel basis — a set such that every vector is a finite combination of basis elements — but for spaces like $\ell^2$ or function spaces this basis is uncountable and non-constructive, requiring the axiom of choice, so you can never write one down. That is useless for physics, which is why Hilbert spaces instead use an orthonormal basis in the topological sense, where "spanning" allows convergent infinite sums. The Fourier basis $\{e^{in x}\}$ is a basis of this second kind, not a Hamel basis. The clean theorems of this article — invariance of dimension, unique finite coordinates — are guaranteed only in the finite-dimensional setting.

Why is the zero vector banned from every independent set?

Because it comes with a built-in dependence. If $\mathbf{0}$ is in your set, then $1\cdot\mathbf{0}=\mathbf{0}$ is already a linear combination that equals the zero vector while using a nonzero coefficient — exactly the nontrivial relation that independence forbids. More intuitively, the zero vector points nowhere and adds no reach, so it is the purest possible redundancy. This is also why any set containing two parallel vectors, or more generally any repeated or scalar-multiple pair, fails independence: a nonzero cancellation is available.

Does the dimension of a space depend on the field of scalars?

Yes, and forgetting this causes real confusion. Dimension is always relative to a chosen field. The complex numbers $\mathbb{C}$ form a one-dimensional vector space over $\mathbb{C}$ (basis $\{1\}$ ), but a two-dimensional space over $\mathbb{R}$ (basis $\{1,i\}$ ) — same set of vectors, different scalar field, different dimension. More generally $\dim_{\mathbb{R}}V=2\,\dim_{\mathbb{C}}V$ for any complex space viewed as real. In quantum mechanics this matters: a qubit's state space is one complex projective dimension but requires two complex (four real) numbers before normalization and phase are removed. Always state the field when you quote a dimension.

How do I actually test independence in practice?

Assemble the vectors as the rows or columns of a matrix and row-reduce. The number of nonzero rows in the echelon form is the rank, which equals the size of the largest independent subset; the vectors are independent exactly when the rank equals their number. For a square case — $n$ vectors in $\mathbb{F}^n$ — the fastest test is the determinant: nonzero means independent (and hence a basis), zero means dependent, as in Worked Example 1. Row reduction also hands you the actual dependence relations, since any column without a pivot is explicitly a combination of the pivot columns.

How do independence and dimension connect to rank and rank–nullity?

They are the same story told about a linear map. The rank of a matrix is the dimension of its column space — the number of independent columns — and the nullity is the dimension of its null space, the solutions of $A\mathbf{x}=\mathbf{0}$ . The rank–nullity theorem, $\operatorname{rank}+\operatorname{nullity}=n$ (the number of columns), is a dimension count: the input space splits into directions the map keeps (rank) and directions it crushes to zero (nullity). So "the columns are independent" is "the rank is full is the nullity is zero is the map is injective" — all one condition, seen through the lens of dimension. This is developed in the rank–nullity article.

Why does the order of a basis matter if it is just a set?

As an abstract certificate of dimension, a basis is just a set and order is irrelevant. But the moment you write coordinates, order becomes essential: the coordinate column $[\mathbf{v}]_{\mathcal{B}}=(v_1,\dots,v_n)^{\mathsf{T}}$ lists the coefficients in a fixed sequence, and permuting the basis permutes the entries. Likewise a matrix representing an operator has its rows and columns indexed by the ordered basis, so reordering the basis conjugates the matrix by a permutation. That is why one speaks of an ordered basis whenever coordinates or matrices are in play, even though independence and spanning themselves never see the order.