The Rank-Nullity Theorem

How many solutions does a linear system have? How does that number relate to the shape of the map itself? This question sits at the heart of linear algebra. Its cleanest answer is the Rank-Nullity Theorem. It says that for a linear map $T$ out of a finite-dimensional vector space $V$ , the dimension of the domain splits exactly into two pieces. One piece is the dimension of the kernel, measuring how much information $T$ destroys. The other is the dimension of the image, measuring how much information survives. This theorem matters because it converts a qualitative question into an arithmetic one. Questions like "is $T$ injective?" or "is $T$ surjective?" become questions about dimensions. Instead of chasing vectors, you count. This is the mechanism behind classical facts like "a system of $n$ equations in $n$ unknowns has a unique solution for every right-hand side if and only if the only solution to the homogeneous system is zero." It is also the engine behind the entire theory of matrix rank. The big picture is this: a linear map $T: V \to W$ cannot lose dimension without a trace. Whatever dimension disappears from the domain into the null space reappears nowhere else. It is simply gone from the range's potential. The theorem quantifies this trade-off perfectly, with no dimension unaccounted for. It is arguably the first deep structural theorem a student meets after basis and dimension themselves. This article gives the precise statement. It includes a complete proof by extending a basis of the null space to a basis of the domain. It offers two fully worked asymmetric examples. It also gives a counterexample showing why finite-dimensionality is essential. Finally, it discusses how the theorem underlies rank computations and the classification of linear systems.

Key takeaways

For a linear map $T: V \to W$ with $\dim V < \infty$ : $\dim V = \dim(\operatorname{null} T) + \dim(\operatorname{range} T)$ .
Proved by extending a basis of $\operatorname{null} T$ to a basis of $V$ , then showing the extra vectors map to a basis of $\operatorname{range} T$ .
Converts injectivity/surjectivity questions into dimension counting.
Requires $\dim V < \infty$ ; fails outright in infinite dimensions (both sides can be "infinite" in incomparable ways).
Immediately implies: $T$ injective iff $\dim(\operatorname{range} T) = \dim V$ ; if $\dim V = \dim W$ , injective iff surjective.
Key idea: dimension lost to the kernel and dimension surviving in the range together account for every dimension of the domain, with none left over.

Where this lives today

In numerical linear algebra, the rank of a matrix (equivalently, of the linear map it represents) governs the solvability structure of $Ax = b$ ; rank-nullity explains why an $m \times n$ system with rank $r$ has a solution space of dimension $n - r$ when solutions exist, directly feeding into the study of least-squares problems and the Singular Value Decomposition. In cryptography and coding theory, linear codes are kernels or images of linear maps over finite fields; the dimension of a code (its "rate") and the dimension of its dual are related by rank-nullity applied to the parity-check map, controlling how many bits of redundancy a code uses. In differential equations, rank-nullity applied to a linear differential operator $L$ acting on a finite-dimensional space of functions (or approximated by a finite-dimensional truncation) explains why the solution space of $Lx = 0$ has a predictable dimension equal to the order of the equation, and constrains when $Lx = f$ is solvable. In quantum mechanics and physics more broadly, linear operators on finite-dimensional state spaces have kernels and images whose dimensions rank-nullity relates; this underlies counting arguments for degenerate eigenspaces and projections. In computer graphics and robotics, linear transformations (projections, rotations composed with degenerate maps) act on coordinate spaces, and rank-nullity explains the dimension of the "null directions" a projection collapses versus the dimension of the image it preserves — essential in inverse kinematics and camera projection models.

Ask the tutor: Can you show me how rank-nullity forces row rank to equal column rank for a matrix?

V

Intuition

Picture $T$ as a machine that takes vectors in $V$ and processes them into vectors in $W$ . Some directions in $V$ are completely destroyed — they get sent to $0$ . These directions sweep out the null space, a subspace of $V$ . Every other direction, roughly speaking, survives and contributes something genuinely new to the output; these directions, together, account for the range. The theorem's proof strategy is to make this "roughly speaking" precise. Pick a basis for the null space — the "destroyed" directions. Extend it (using the basis extension theorem, a standard fact about finite-dimensional vector spaces) to a basis of the whole domain $V$ . The new, extra basis vectors are exactly the ones that survive, and the claim is that their images form a basis of the range. Once you see this, dimension counting does the rest: the number of vectors used for the null space, plus the number of extra vectors, equals $\dim V$ ; and the extra vectors correspond bijectively to a basis of the range. Think of it as a conservation law. The "mass" of the domain, measured in dimension, cannot vanish — it must show up either as nullity (destroyed) or as rank (surviving in the image). Nothing else can happen to it.

Figure

Domain dimension splits into null-space dimension and rank

V

\operatorname{null} T

V, W

T

Worked examples

Example 1: A map from $\mathbf{R}^4$ to $\mathbf{R}^3$ . Define $T: \mathbf{R}^4 \to \mathbf{R}^3$ by

T(x_1, x_2, x_3, x_4) = (x_1 + 2x_2 - x_3 + x_4,\; 2x_1 + 4x_2 - 2x_3 + 3x_4,\; x_1 + 2x_2 - x_3 + 2x_4).

Finding the null space. We solve $T(x) = 0$ :

\begin{aligned} x_1 + 2x_2 - x_3 + x_4 &= 0 \\ 2x_1 + 4x_2 - 2x_3 + 3x_4 &= 0 \\ x_1 + 2x_2 - x_3 + 2x_4 &= 0 \end{aligned}

Subtract twice the first equation from the second: $(2x_1 + 4x_2 - 2x_3 + 3x_4) - 2(x_1+2x_2-x_3+x_4) = x_4 = 0$ . Subtract the first equation from the third: $(x_1+2x_2-x_3+2x_4) - (x_1+2x_2-x_3+x_4) = x_4 = 0$ (consistent, no new information). So $x_4 = 0$ , and the first equation becomes $x_1 + 2x_2 - x_3 = 0$ , i.e. $x_1 = -2x_2 + x_3$ . Free variables: $x_2, x_3$ . So

(x_1,x_2,x_3,x_4) = x_2(-2,1,0,0) + x_3(1,0,1,0).

Thus $\operatorname{null} T = \operatorname{span}\{(-2,1,0,0), (1,0,1,0)\}$ . These two vectors are not scalar multiples of each other (their second and third coordinates already show independence: the first has a nonzero second coordinate and zero third, the second has zero second coordinate and nonzero third), so they are linearly independent, giving

\dim(\operatorname{null} T) = 2.

Applying the theorem. $\dim V = 4$ , so the theorem predicts

\dim(\operatorname{range} T) = 4 - 2 = 2.

Direct check. $\operatorname{range} T$ is spanned by the images of the standard basis vectors:

T(e_1) = (1,2,1),\ T(e_2) = (2,4,2),\ T(e_3) = (-1,-2,-1),\ T(e_4) = (1,3,2).

Note $T(e_2) = 2\,T(e_1)$ and $T(e_3) = -T(e_1)$ , consistent with $e_2 - 2e_1 \cdot(\text{part})$ ... more directly, this redundancy reflects that $(-2,1,0,0)$ and $(1,0,1,0)$ (built from $e_1,e_2,e_3$ ) lie in the null space. The range is spanned by $T(e_1) = (1,2,1)$ and $T(e_4) = (1,3,2)$ , since $T(e_2), T(e_3)$ are multiples of $T(e_1)$ . Check independence: is $(1,3,2) = \lambda(1,2,1)$ for some $\lambda$ ? That would need $\lambda = 1$ from the first coordinate and $\lambda = 3/2$ from the second — contradiction. So $T(e_1), T(e_4)$ are independent, confirming

\dim(\operatorname{range} T) = 2,

exactly matching $4 - 2 = 2$ predicted by the theorem.

Example 2: A map from $\mathbf{R}^3$ to $\mathbf{R}^5$ (rank-nullity with a "wide" codomain). Define $T: \mathbf{R}^3 \to \mathbf{R}^5$ by

T(x,y,z) = (x - y,\; 2x + z,\; -y - z,\; x + z,\; 3x - 2y + z).

Null space. Solve $T(x,y,z) = 0$ :

\begin{aligned} x - y &= 0 \\ 2x + z &= 0 \\ -y - z &= 0 \\ x + z &= 0 \\ 3x - 2y + z &= 0 \end{aligned}

From equation 1: $y = x$ . From equation 2: $z = -2x$ . Check equation 3: $-y - z = -x - (-2x) = x$ ; for this to be $0$ we need $x = 0$ . Then $y = 0, z = 0$ . Check equations 4,5 with $x=y=z=0$ : both hold trivially. So the only solution is $(0,0,0)$ , hence

\operatorname{null} T = \{0\}, \qquad \dim(\operatorname{null} T) = 0.

Applying the theorem. $\dim V = 3$ , so

\dim(\operatorname{range} T) = 3 - 0 = 3.

Direct check. Since $\operatorname{null} T = \{0\}$ , $T$ is injective (Corollary above, direction already built into the definition: injective iff trivial kernel). An injective linear map sends a basis of the domain to a linearly independent list in the codomain (if $\sum c_i Tv_i = 0$ then $T(\sum c_i v_i) = 0$ forces $\sum c_i v_i = 0$ by injectivity, hence all $c_i = 0$ by independence of $v_i$ in $V$ ). Applying $T$ to the standard basis of $\mathbf{R}^3$ :

T(e_1) = (1,2,0,1,3), \quad T(e_2) = (-1,0,-1,0,-2), \quad T(e_3) = (0,1,-1,1,1).

By injectivity these three vectors in $\mathbf{R}^5$ are linearly independent, so they span a $3$ -dimensional subspace of $\mathbf{R}^5$ equal to $\operatorname{range} T$ . This matches $\dim(\operatorname{range} T) = 3$ exactly as predicted, and also illustrates that $T$ is injective but (since $3 < 5$ ) cannot be surjective onto all of $\mathbf{R}^5$ .

V

Scope, edge cases & common traps

Situation	What happens
$\dim V < \infty$ , $W$ arbitrary (possibly infinite-dimensional)	Theorem holds; $\operatorname{range} T$ is automatically finite-dimensional, with $\dim(\operatorname{range} T) \le \dim V$ .
$\dim V = \infty$	Theorem's numerical identity is not applicable; nullity or rank (or both) may be infinite, and no clean "subtraction" identity survives, as shown above.
$T$ is the zero map	$\operatorname{null} T = V$ , $\operatorname{range} T = \{0\}$ ; identity reads $\dim V = \dim V + 0$ . Degenerate but consistent.
$T$ is an isomorphism ( $\dim V = \dim W$ , bijective)	$\operatorname{null} T = \{0\}$ , $\operatorname{range} T = W$ ; identity reads $\dim V = 0 + \dim W$ .
$\dim V > \dim W$	$T$ cannot be injective, since $\dim(\operatorname{range} T) \le \dim W < \dim V$ forces $\dim(\operatorname{null} T) > 0$ by the theorem.
$\dim V < \dim W$	$T$ cannot be surjective, since $\dim(\operatorname{range} T) \le \dim V < \dim W$ .

Common trapThe classic error: assuming the theorem for infinite-dimensional domains, or reversing the injective/surjective corollary's hypothesis

Two mistakes recur constantly. First, students sometimes try to apply $\dim V = \operatorname{nullity} + \operatorname{rank}$ to maps between infinite-dimensional spaces without checking $\dim V < \infty$ ; as the counterexample above shows, this can produce meaningless or false conclusions. Second, students misuse the corollary " $T$ injective iff surjective" without first checking $\dim V = \dim W$ : this equivalence is false in general when the dimensions differ (e.g. the inclusion $\mathbf{R}^2 \hookrightarrow \mathbf{R}^3$ , $(x,y)\mapsto(x,y,0)$ , is injective but not surjective). The quantifier order matters: the hypothesis $\dim V = \dim W$ must be checked before invoking the equivalence, not assumed as a consequence of it.

The mathematician behind it

The Rank-Nullity Theorem, in the language of abstract vector spaces, crystallized as part of the broader 20th-century axiomatization of linear algebra, but its essential content — the dimension-counting relationship between a linear system's solution space and its image — traces back through 19th-century work on linear equations, determinants, and matrices.

Origins · 1 minGeorg Frobenius and how this was discovered

Georg Frobenius

1849–1917 · German mathematician, rank of matrices and representation theory

Frobenius made foundational contributions to the theory of matrices, including systematic study of the rank of a matrix and its role in the solvability of linear systems, work that runs in parallel with the modern rank-nullity relationship. His broader legacy includes deep contributions to group representation theory and the theory of bilinear forms, areas in which dimension-counting arguments of exactly this rank-nullity flavor recur constantly.

History

- 1878: Frobenius develops systematic matrix rank theory, connecting the rank of the coefficient matrix to solvability of linear systems.

- 1888: Giuseppe Peano gives an early axiomatic treatment of vector spaces, providing the abstract setting in which "dimension" is properly defined.

- 1920s–1930s: Hermann Weyl, Emmy Noether, and others promote the abstract, basis-free treatment of linear maps between vector spaces, in which rank-nullity becomes a theorem about arbitrary finite-dimensional

V, W

rather than only matrices.

- 1930s: The theorem enters standard textbook form (kernel/image language) as part of the modern algebraic reformulation of linear algebra.

- Mid-20th century onward: Rank-nullity becomes a standard early theorem in university linear algebra curricula worldwide, underlying the modern treatment of Rank of a Matrix and systems of linear equations.

Does the Rank-Nullity Theorem apply to linear maps between infinite-dimensional spaces?

Not in its numerical form. The theorem as stated requires $\dim V < \infty$ , because the proof relies on extending a basis of the null space to a finite basis of $V$ and then counting vectors. If $V$ is infinite-dimensional, there may be no finite basis to extend, and the quantities $\dim(\operatorname{null} T)$ , $\dim(\operatorname{range} T)$ can individually be finite or infinite in combinations that admit no clean additive identity (see the differentiation-operator counterexample above, where the null space is finite-dimensional but the range is infinite-dimensional). Some infinite-dimensional analogues exist using cardinal arithmetic, but they require substantially more care and are outside the elementary theorem's scope.

Why must the extending vectors

v_1, \ldots, v_n

come from extending a basis of

\operatorname{null} T

, rather than any complementary spanning set?

The proof needs two separate facts about $Tv_1,\dots,Tv_n$ : that they span $\operatorname{range} T$ , and that they are linearly independent. The independence argument crucially uses that $u_1,\dots,u_m,v_1,\dots,v_n$ together form a basis of $V$ — this is what lets us conclude, from a dependency among the $Tv_i$ 's, that a certain combination of the $v_i$ 's lying in $\operatorname{null} T$ must in fact be a combination of the $u_i$ 's, and then invoke linear independence of the full list to force all coefficients to zero. An arbitrary complementary spanning set without this "basis of $V$ " property would not support that final independence step.

Is the converse of the Rank-Nullity Theorem true — does any pair of numbers satisfying the identity correspond to some linear map?

Yes, in the following precise sense: given finite-dimensional $V$ and $W$ over $\mathbf{F}$ and any nonnegative integers $m, n$ with $m + n = \dim V$ and $n \le \dim W$ , one can construct a linear map $T: V \to W$ with $\dim(\operatorname{null} T) = m$ and $\dim(\operatorname{range} T) = n$ . Sketch: choose a basis $u_1,\dots,u_m,v_1,\dots,v_n$ of $V$ and linearly independent vectors $w_1,\dots,w_n$ in $W$ (possible since $n \le \dim W$ ); define $T$ on the basis by $Tu_i = 0$ , $Tv_j = w_j$ , and extend linearly. This is not automatically implied by the theorem itself but is a natural companion construction, provable directly and not requiring any deeper machinery.

How does rank-nullity relate to solving the linear system

Ax = b

View $A$ as a linear map $T: \mathbf{F}^n \to \mathbf{F}^m$ via $T(x) = Ax$ . The homogeneous system $Ax = 0$ has solution set exactly $\operatorname{null} T$ , so its solution space has dimension $n - \operatorname{rank}(A)$ by the theorem. If $Ax = b$ has at least one solution $x_0$ , every solution has the form $x_0 + z$ with $z \in \operatorname{null} T$ (since $A(x_0+z) = Ax_0 + Az = b + 0 = b$ , and conversely any two solutions differ by an element of $\operatorname{null} T$ ). So the full solution set, when nonempty, is an affine translate of $\operatorname{null} T$ , whose dimension the theorem pins down exactly as $n - \operatorname{rank}(A)$ .

Does the theorem tell you whether

Ax = b

has a solution at all?

No — rank-nullity describes the dimension of null space and range but says nothing by itself about whether a specific $b$ lies in $\operatorname{range} T$ . Solvability of $Ax=b$ is governed separately by whether $b \in \operatorname{range}(A)$ , equivalently (for augmented-matrix criteria) whether $\operatorname{rank}[A \mid b] = \operatorname{rank}(A)$ . Rank-nullity instead tells you, once you know solvability, exactly how large the solution set is: an affine subspace of dimension $n - \operatorname{rank}(A)$ , as described above.

How does this theorem foreshadow the study of matrix rank?

For a matrix $A$ viewed as the linear map $x \mapsto Ax$ , the rank of $T$ in the sense above (dimension of $\operatorname{range} T$ , the column space of $A$ ) coincides with the usual matrix rank. Rank-nullity then becomes the identity relating the number of columns $n$ to the dimension of the null space (solution space of $Ax=0$ ) and the column rank. This sets up the deeper fact, proved separately, that row rank equals column rank for any matrix — a cornerstone result explored under Rank of a Matrix and central to the broader study of Consequences of Rank-Nullity for Finite Dimensions.