Every linear map carries two intrinsic subspaces that reveal its structure: the kernel (what gets annihilated) and the image (what gets hit). These are the linear-algebraic analogues of concepts we have seen before: the kernel of a group homomorphism and the image of a function. The relationship between kernel and image is governed by the Rank–Nullity Theorem.
Note. Kernel and image measure two different things. The kernel measures the failure of injectivity—how much information the map destroys. The image measures the failure of surjectivity—how much of the codomain the map reaches. The Rank–Nullity Theorem says these two failures are complementary: the more a linear map kills, the less it can produce, and vice versa.
Kernel and image do not determine the map. Every automorphism of has kernel and image ; for instance and (when in ) have the same kernel and the same image. What is true is this: if is a complement of in (one exists when is finite-dimensional, by Proposition 4.7.11), so that , then for and , and restricts to an isomorphism (its kernel is ). So is determined by its kernel together with this isomorphism, and knowing the kernel and the image leaves the isomorphism undetermined.
6.1.1 The Kernel
Definition 6.1.1 (Kernel of a linear map). Let be a linear map. The kernel of is This is the set of all vectors that maps to zero.
Note. The term “kernel” comes from algebra (group homomorphisms), while “null space” is the synonym more common in applied linear algebra and matrix theory; Axler writes for . In Chapter 7 we will meet the generalized kernel of an operator , the set of vectors killed by some power of . It contains and can be strictly larger.
Example 6.1.2 (Examples of kernels).
- The zero map has .
- The identity map has .
- Let be defined by . Then , a plane through the origin.
- Let be differentiation, . Then is the set of constant polynomials.
- The trace of Chapter 5 is linear, and is the space of matrices of trace zero.
Proposition 6.1.3 (The kernel is a subspace). Let be a linear map. Then is a subspace of .
Proof. We verify the subspace criteria:
- Contains zero: Since (every linear map preserves zero), we have .
- Closed under addition: If , then , so .
- Closed under scalar multiplication: If and , then , so .
∎
Note. Compare this with group theory: the kernel of a group homomorphism is a normal subgroup of . For linear maps, every subspace is “normal” in the appropriate sense—vector spaces are abelian groups under addition, so there is no distinction between left and right cosets.
The kernel provides a clean characterization of injectivity:
Theorem 6.1.4 (Injectivity via the kernel). Let be a linear map. Then
Proof. Suppose is injective. Let . Then . Since is injective, . Thus , and since always, we have .
Suppose . Let with . Then so . Thus , i.e., . Hence is injective. ∎
Note. The kernel does more than detect injectivity—it measures the failure of injectivity in a precise sense. Two vectors have the same image under if and only if . Thus the fibers of (preimages of points of ) are exactly the cosets . The map is injective precisely when these cosets are singletons, i.e., when . This is why we can “mod out by the kernel” to obtain an injective map: the quotient space , constructed in Section 7.1, collapses each fiber to a point.
6.1.2 The Image
Definition 6.1.5 (Image of a linear map). Let be a linear map. The image (or range) of is This is the set of all vectors that “hits.”
Example 6.1.6 (Examples of images).
- The zero map has .
- The identity map has .
- Let be the projection . Then .
- For , let be differentiation. Then : every polynomial of degree is the derivative of a polynomial of degree .
- Let define , . Since , where are the columns of , the image is the column space of , the span of its columns.
Proposition 6.1.7 (The image is a subspace). Let be a linear map. Then is a subspace of .
Proof. We verify the subspace criteria:
- Contains zero: , so .
- Closed under addition: If , then and for some . Thus .
- Closed under scalar multiplication: If and , then for some , so .
∎
Corollary 6.1.8. An injective linear map embeds as a subspace of : it is an isomorphism from onto the subspace of .
Proof. Regarded as a map , is linear and surjective, and it is injective by hypothesis. ∎
Definition 6.1.9 (Surjectivity). A linear map is surjective (or onto) if , i.e., every element of is hit by some element of .
Proposition 6.1.10 (The image of a spanning set). Let be a linear map and let be a subset of that spans . Then In particular, is surjective if and only if spans , and if is finite-dimensional, then so is .
Proof. Since is a subspace containing , it contains . Conversely, every is a linear combination of elements of , so . If is finite-dimensional, take to be a finite basis; then is spanned by finitely many vectors. ∎
6.1.3 The Rank–Nullity Theorem
We now come to the central result connecting kernel and image. This theorem quantifies the fundamental trade-off in linear algebra: the more a map “kills,” the less it can “produce.”
Definition 6.1.11 (Rank and nullity). Let be a linear map with finite-dimensional. Then is finite-dimensional (it is a subspace of ; Proposition 4.7.10), and so is (Proposition 6.1.10). The rank of is and the nullity of is For a matrix , we define and to be the rank and nullity of .
Theorem 6.1.12 (The Rank–Nullity Theorem). Let be a linear map with finite-dimensional. Then Equivalently: . More precisely, if is a basis of and is a basis of , then is a basis of .
Proof. The kernel is a subspace of the finite-dimensional space , so it has a finite basis (Proposition 4.7.10), and every linearly independent list extends to a basis of (Lemma 4.7.4). So bases as in the statement exist, and . It suffices to prove the last sentence.
Claim: is a basis for .
Spanning: Let , say for some . Write Applying : Since each , we have , so Thus spans .
Linear independence: Suppose Then , so . Since is a basis for , we can write for some scalars . Rearranging: Since is linearly independent (it is a basis of ), all coefficients are zero. In particular, .
Thus is a basis for , so . We conclude: ∎
Note. The Rank–Nullity Theorem is the linear-algebra counterpart of the First Isomorphism Theorem for groups (Theorem 3.3.3), which says for a homomorphism . In Section 7.1 we construct the quotient space and show ; Rank–Nullity is the statement that this isomorphism is compatible with counting dimensions, . This is a way to remember the theorem and is not a second proof of it: the formula of Chapter 7 is itself deduced from Rank–Nullity.
The proof of the Rank–Nullity Theorem produces bases in which a linear map looks as simple as possible.
Theorem 6.1.13 (Normal form of a linear map). Let be a linear map between finite-dimensional vector spaces, with , , and . Then there exist a basis of and a basis of such that where is the identity matrix and the zeros are zero blocks of the appropriate sizes.
Proof. Choose a basis of , and extend it by vectors to a basis of . By Theorem 6.1.12, the vectors form a basis of ; extend them to a basis of . Take and . For , , so the -th column of the matrix is the -th standard basis vector of ; the remaining columns are zero because . ∎
Note. In this normal form the first vectors of span and the last vectors of span . The only information about left in the matrix is the number : once we may choose bases of and independently, the rank is the only invariant of a linear map.
Corollary 6.1.14 (Matrices up to change of bases). Let have rank . Then there are invertible matrices and with Consequently, two matrices satisfy for some invertible and if and only if .
Proof. Apply Theorem 6.1.13 to , and let and be the matrices whose columns are the vectors of and of . By the change-of-basis formula (Theorem 5.3.3), the matrix of in the bases and is .
If , then because is surjective, and maps isomorphically onto ; hence . Conversely, if both matrices have rank , then for suitable invertible , and . ∎
6.1.5 Consequences of the Rank–Nullity Theorem
The Rank–Nullity Theorem has powerful implications for understanding linear maps.
Corollary 6.1.15 (Dimension constraints on injectivity). Let be a linear map between finite-dimensional spaces with and .
- If , then cannot be injective.
- If , then cannot be surjective.
Proof. (a) If were injective, then , so by Rank–Nullity: But , so . Contradiction.
(b) If were surjective, then , so by Rank–Nullity: But dimensions are non-negative. Contradiction. ∎
Corollary 6.1.16 (Equal dimensions: bijectivity criteria). Let be a linear map with finite. The following are equivalent:
- is injective.
- is surjective.
- is bijective (an isomorphism).
- .
- .
- .
Proof. By Rank–Nullity, . A subspace of of dimension is all of , so if and only if . Using Theorem 6.1.4:
Thus injectivity, surjectivity, and bijectivity are equivalent when dimensions match. ∎
Note. For linear maps between equal-dimensional spaces, “one-to-one” and “onto” are the same condition. This fails for general functions and even for linear maps between infinite-dimensional spaces. Consider the right shift defined by . This is injective (its kernel is trivial) but not surjective (the sequence is not in the image). Conversely, the left shift is surjective but not injective. In infinite dimensions there is “room” for one without the other. The finite-dimensional equivalence is the linear analogue of the pigeonhole principle.
Corollary 6.1.17 (Existence of left and right inverses). Let be a linear map between finite-dimensional spaces.
- has a left inverse (a linear with ) if and only if is injective.
- has a right inverse (a linear with ) if and only if is surjective.
Proof. (a) If and , then , so is injective. Conversely, let be injective and let be a basis of . The vectors are linearly independent: if , then , so and all . Extend them to a basis of , and define on this basis by and (a linear map may be prescribed arbitrarily on a basis, Proposition 5.1.2). Then agrees with on a basis, so .
(b) If , then every equals , so is surjective. Conversely, let be surjective and let be a basis of . Choose with , and let be the linear map with . Then agrees with on a basis, so . ∎
6.1.6 Matrix Interpretation
For a matrix , the associated linear map is given by . The Rank–Nullity Theorem becomes:
Proposition 6.1.18 (Rank–Nullity for matrices). For :
- is the null space of .
- is the column space of (the span of the columns).
- (the number of columns).
Note. Since the columns of span the column space and every spanning list contains a basis (Lemma 4.7.2), the rank of is the largest number of linearly independent columns of . It is also the largest number of linearly independent rows; this is Corollary 6.2.17 below, “row rank equals column rank.” The null space has dimension .
Looking ahead to Chapter 7: for a linear operator and a scalar , the eigenspace is the kernel of a shifted operator. Its dimension (the geometric multiplicity of ) measures “how degenerate” is, and the gap between geometric and algebraic multiplicities is what makes the Jordan normal form necessary.
Example 6.1.19 (Computing kernel and image). Let over .
Finding : Solve . Row operations do not change the solutions of , so we row reduce: Back-substituting: , . So
Finding : The echelon form has its pivots in columns 1 and 2. Since row operations preserve the linear relations among the columns (a vector satisfies if and only if it satisfies for the reduced matrix ), columns 1 and 2 of are linearly independent, and column 3 is a combination of them: . Hence
Verification: . ✓
Question 6.1.20.
- Let be the differentiation operator. Find , , , and . Verify the Rank–Nullity Theorem.
- Let be defined by . Determine whether is injective, surjective, or neither.
- Prove that if is a linear operator with , then .
- Prove that a linear map is injective if and only if for every linearly independent list in , its image under is linearly independent in .
Show solution
(a) The space has basis , so .
Kernel: if and only if is a constant polynomial. So , the constant polynomials. Thus .
Image: Differentiating: , , , . By Proposition 6.1.10, these span the image, so and .
Verification: . ✓
(b) Injectivity: Find . Solve : This gives . So .
Since , is not injective.
Surjectivity: We have , so . Thus , and is surjective.
(c) We show and .
Sum equals : Let . Write . Note that . Also: so . Thus .
Intersection is trivial: Let . Then and for some . So:
Thus .
(d) () Suppose is injective and is linearly independent. Suppose . Then , so . Thus , and by linear independence of the , all . Hence is linearly independent.
() Suppose maps every linearly independent list to a linearly independent list. If , then the list is linearly independent, so is linearly independent, meaning . Thus , so is injective.