S. D. V. Stephens

notes / Studies in Algebra and Group Theory / chapter 3: Linear Maps and Operators / From my book

B6.2 Duality

The dual space of a vector space consists of all linear maps from to the base field —these are called linear functionals. While this might seem like a minor variation on Hom-spaces, duality reveals useful structure: every basis has a “dual basis,” every linear map has a “dual map,” and the matrix of the dual map is the transpose of the original. In more advanced contexts (differential geometry, functional analysis, representation theory), duality becomes indispensable.

6.2.1 Linear Functionals and the Dual Space

Definition 6.2.1 (Linear functional). A linear functional on a vector space over a field is a linear map . In other words, a linear functional is an element of .

Example 6.2.2 (Examples of linear functionals).

  1. Coordinate projections: On , define . Each is a linear functional.
  2. Weighted sums: Fix . Define . This is a linear functional on , and every linear functional on has this form, with : indeed .
  3. Evaluation: On , define for fixed . This is a linear functional.
  4. Integration: On , define . Linearity of the integral makes this a linear functional.
  5. Trace: On , the trace is a linear functional.

Definition 6.2.3 (Dual space). The dual space of , denoted , is the vector space of all linear functionals on : This is a vector space under pointwise addition and scalar multiplication: and .

Note. Axler writes for the dual space and for the annihilator defined below. In this book the star is reserved for dual spaces; for the dual of a map we write , because will denote the adjoint of an operator on an inner product space (Chapter 9).

Proposition 6.2.4 (Dimension of the dual space). If is finite-dimensional with , then . Thus as vector spaces (non-canonically).

Proof. By the formula (Corollary 5.2.8): ∎

Note. The isomorphism is non-canonical—it depends on a choice of basis. There is no “natural” way to identify vectors with linear functionals without extra structure. However, the double dual does have a canonical isomorphism with , given by evaluation: (Theorem 6.2).

6.2.2 The Dual Basis

Given a basis of , there is an associated basis of .

Definition 6.2.5 (Dual basis). Let be a basis of . The dual basis is the list of linear functionals in defined by (A linear map is determined by its values on a basis, and any values may be prescribed; Proposition 5.1.2.) In other words, “picks out” the -th coordinate: if , then .

Proposition 6.2.6 (The dual basis is a basis). The dual basis is a basis of .

Proof. Linear independence: Suppose (the zero functional). Evaluating at : So all .

Spanning: Since we have linearly independent vectors in and , they form a basis. ∎

Proposition 6.2.7 (Expansions in a basis and its dual basis). Let be a basis of with dual basis . For any and any : That is, is the -th coordinate of in the basis , and is the -th coordinate of in the dual basis.

Proof. Write . Applying : Substituting back gives the first formula. For the second, both sides are linear functionals, and at the right side takes the value . Two linear maps that agree on a basis are equal. ∎

Example 6.2.8 (Dual basis of the standard basis). Let be the standard basis of . The dual basis consists of the coordinate projections: If we write elements of as column vectors and represent the functional by the row vector , so that applying a functional to a vector is the matrix product (row)(column), then is the row vector with in position .

Note. Each depends on the whole basis: changing the other basis vectors can change even when stays the same. There is no such thing as “the dual of a vector.” In , the basis gives , while the basis , which has the same first vector, gives the first dual functional . This is another way of seeing that the isomorphism depends on choices.

6.2.3 The Dual Map

Just as a linear map acts on Hom-spaces by composition, it induces a map on dual spaces—but in the opposite direction.

Definition 6.2.9 (Dual map (transpose)). Let be a linear map. The dual map (or transpose) of is the map defined by for each . In other words, is the linear functional on given by .

Note. The dual map goes “backwards”: yields , because we are precomposing with . If you think of linear functionals as “measurements,” then is the measurement you get by first applying , then measuring with .

Proposition 6.2.10 (The dual map is linear). For , the dual map is linear.

Proof. For and : ∎

Proposition 6.2.11 (Algebraic properties of the dual map). Let be vector spaces over .

  1. For , .
  2. For and , .
  3. For and , (the order reverses).
  4. .

Proof. (c) is the key property. For : For (a), , because is additive; (b) is the same computation with ; and (d) holds because . ∎

Note. Properties (c) and (d) say that , reverses the order of composition and preserves identities; in the language of Chapter 8, it is a contravariant functor (Definition 8.2.5 and Example 8.2.6). Compare with the transpose of matrices: . This is not a coincidence—the matrix of the dual map is the transpose.

6.2.4 Matrix of the Dual Map

The connection between dual maps and matrix transposes explains why the transpose operation is natural.

Theorem 6.2.12 (Matrix of the dual map is the transpose). Let be a linear map between finite-dimensional spaces. Fix bases of and of , with dual bases and . If is the matrix of , then the matrix of with respect to the dual bases is the transpose:

Proof. We need to show that for each .

By definition of the matrix , we have . Now compute: By Proposition 6.2.7, . So the -th column of is the -th row of . Hence . ∎

Note. This theorem explains why the transpose is natural: it is the coordinate representation of the dual map. The reversal of dimensions ( becomes ) reflects the reversal of arrows. In terms of row vectors: if is represented by the row vector and by the column vector , then , so acts on row vectors by ; writing row vectors as columns turns this into .

6.2.5 Annihilators

Definition 6.2.13 (Annihilator). Let be a subspace of . The annihilator of is the set of linear functionals on that vanish on .

The annihilator is a subspace of : the zero functional vanishes on , and if and vanish on , so does . Directly from the definition, , , and implies : bigger subspaces have smaller annihilators.

Theorem 6.2.14 (Dimension of the annihilator). Let be finite-dimensional and a subspace. The restriction map is linear and surjective, and its kernel is . Thus is a short exact sequence (Definition 3.5.1): the inclusion of is injective, is surjective, and the image of each map is the kernel of the next. Consequently

Proof. Linearity of is clear, and means exactly that vanishes on , so . For surjectivity, let . Choose a basis of and extend it to a basis of . Let be the functional with for and for . Then and agree on a basis of , so .

By Rank–Nullity applied to , Since and , we get . ∎

Note. The annihilator gives an inclusion-reversing correspondence between subspaces of and subspaces of : a subspace of dimension in an -dimensional has an annihilator of dimension . In Section 7.1, once quotient spaces are available, we will see that and (Theorem 7.1.10).

6.2.6 Kernel and Image of the Dual Map

The dual map interchanges the roles of kernel and image in a precise way.

Theorem 6.2.15 (Kernel and image of the dual map). Let be a linear map.

  1. .
  2. If and are finite-dimensional, then , and .

Proof. (a) For :

(b) First, : if , then for every . Next we compare dimensions. Let , , and . By Rank–Nullity for , by (a), and by Theorem 6.2.14, On the other hand, , by Theorem 6.2.14 and Rank–Nullity for . A subspace of dimension contained in a space of dimension is the whole space, so . ∎

Corollary 6.2.16 (Injectivity and surjectivity of the dual map). Let be a linear map between finite-dimensional spaces. Then

  1. is surjective if and only if is injective;
  2. is injective if and only if is surjective.

Proof. (a) By Theorem 6.2.15(a) and Theorem 6.2.14, , which is exactly when , i.e., when is surjective.

(b) By Theorem 6.2.15(b), . This equals , i.e., is surjective, exactly when . ∎

Note. The implication “ surjective injective” holds without any finiteness assumption: if is surjective and , then vanishes on , so . The other implications use dimension counts.

Corollary 6.2.17 (Row rank equals column rank). For any matrix , . In words: the largest number of linearly independent rows of equals the largest number of linearly independent columns.

Proof. The matrix of in the standard bases is , so by Theorem 6.2.12 the matrix of in the dual bases is . The coordinate isomorphisms and given by the dual bases carry to , and isomorphisms preserve the dimension of images. Hence , using Theorem 6.2.15(b). The rows of are the columns of , which gives the second sentence. ∎

6.2.7 The Double Dual

We have seen that for finite-dimensional , but this isomorphism required a choice of basis. The double dual enjoys a more canonical relationship with .

Definition 6.2.18 (Evaluation map). The evaluation map is defined by for and . That is, is the linear functional on that “evaluates at .” (It is linear in because the operations on are pointwise: .)

Theorem 6.2.19 (The double dual). The evaluation map is linear. If is finite-dimensional, then is an isomorphism.

Proof. Linearity: For , , and :

Injectivity (for finite-dimensional ): Suppose , meaning for all . If , extend to a basis of ; the first dual basis functional satisfies , a contradiction. So .

Isomorphism: For finite-dimensional , . Since is injective, it is an isomorphism (Corollary 6.1.16). ∎

Note. The point is that involves no choices. Given a basis of , its dual basis of , and the dual basis of , we have , so . The basis-dependent isomorphisms and compose to an isomorphism that does not depend on the basis.

Moreover, is compatible with every linear map : writing and for the evaluation maps of and , since for both sides send to the functional . Chapter 8 expresses this by saying that is a natural transformation (Definition 8.2.9).

Note. For infinite-dimensional , the situation is strikingly different. Suppose has a basis indexed by an infinite set (every vector space has a basis, a consequence of Zorn’s lemma). Every element of is a finite linear combination (with only finitely many ). However, elements of can be arbitrary on the basis: given any function , the map is a well-defined linear functional. Thus , the full product, which is much larger than the direct sum ; in particular the coordinate functionals do not span . The evaluation map is still injective (for , some coordinate functional has ), but it is not surjective.

Question 6.2.20.

  1. Let with standard basis . Write down the dual basis explicitly as functions .
  2. Let be defined by . Find the matrix of with respect to the standard bases, then find the matrix of with respect to the dual bases.
  3. Let . Find a basis of and check the dimension formula of Theorem 6.2.14.
  4. Let be finite-dimensional and a subspace. Prove that , where is the evaluation map. (So under the identification of with , taking the annihilator twice returns .)
Show solution

(a) The dual basis consists of the coordinate projections: As row vectors these are , , .

(b) The matrix of is: The matrix of is the transpose: Explicitly, if is the dual basis of the standard basis of and is that of , then is , and similarly for : The coefficients form the columns of , as the theorem predicts.

(c) A functional lies in if and only if it vanishes on the two spanning vectors: and . So and , and is spanned by the single functional The two spanning vectors of are linearly independent, so , and indeed .

(d) Here and . If and , then ; hence . Let . Applying Theorem 6.2.14 twice (to and to ), Since is injective, as well. A subspace contained in another of the same finite dimension equals it, so .