Matrices were introduced as arrays that organize linear systems and then as objects with their own algebra. This unit explains why the multiplication rule for matrices is the right one: a matrix is the same thing as a function that respects addition and scalar multiplication. A function of this kind is called a linear transformation. The unit uses the definition of a vector space (6.1), subspaces (6.2), basis and dimension (6.5), and matrix multiplication (3.2), and it prepares the rank-nullity theorem in the next unit.
Definition and first consequences
Definition 6.9.1
(Linear transformation)
Let and be real vector spaces. A function is a linear transformation (or linear map) if for all and all real ,
The two conditions together are equivalent to the single condition .
Proposition 6.9.2
(Linear maps preserve zero, negatives and linear combinations)
If is linear, then , , and for all vectors and scalars .
Proof
(Zero and negatives from scalar multiples, combinations by induction)
. Next, . The last formula follows from the two defining conditions by induction on .
The condition gives a quick test for failure: the function is not linear because . But is only necessary.
Counterexample mode
Sending zero to zero is not enough
Claim: if , then is linear. Take . Then , but
so homogeneity fails and is not linear. The repaired statement is only the converse direction: a linear map sends to , so rules a function out, while rules nothing in.
Every linear map from to is a matrix
For an real matrix , the function , , is linear by the distributive law . The converse is the main structural fact.
Theorem 6.9.3
(Standard matrix of a linear map)
Let be linear and let be the standard basis of . Let be the matrix whose -th column is . Then for all , and is the only matrix with this property.
Proof
(Expand in the standard basis, then read off the columns)
Write . By the previous proposition,
because is by definition the combination of the columns of with coefficients . If is another matrix with for all , then is the -th column of , so .
Proof X-ray
Where linearity and the basis enter
The proof has an existence half and a uniqueness half, and each uses a different property. Existence needs two facts: every vector is a combination of (the spanning property of the basis), and passes through combinations (linearity, in the form of the previous proposition). The first fact turns an arbitrary input into finitely many known ones, ; the second turns the unknown value into the same combination of the known values . No other property of is used, and no choice is made: the columns are forced. Uniqueness is a separate observation, that the product picks out column of , so two matrices that agree on have the same columns.
The closing step is the identity , which is just the definition of the matrix–vector product as a combination of columns. This is why the multiplication rule is built that way. Without linearity the expansion step fails: for the non-linear of the previous example, the values and would force , but in fact .
The matrix is the standard matrix of and is written . The theorem says that a linear map is determined by its values on a basis, and that these values can be chosen freely.
Quick check
(Find the standard matrix of .)
Compute and and use them as columns.
Solution · Answer
and , so . Check: .
Concept lensGeometric
The columns as the image of the unit square
In the plane, the standard matrix has a picture. The unit square has sides ; its image under is the parallelogram with sides , the two columns of . Linearity means that grid lines stay straight and equally spaced, so the whole grid is carried to the grid of that parallelogram. For , the square is sent to the parallelogram with sides and , and the point goes to .
Composition, kernel and range
Theorem 6.9.4
(Composition corresponds to matrix product)
If and are linear, then is linear and .
Proof
(Associativity plus uniqueness of the standard matrix)
Linearity of is immediate from the definitions. For the matrix, by associativity of matrix multiplication, and uniqueness in the standard-matrix theorem gives the claim.
This is the reason matrix multiplication is defined as it is, and also the reason in general: composing functions in different orders gives different functions.
Definition 6.9.5
(Kernel and range)
For a linear map the kernel and the range are
Theorem 6.9.6
(Kernel, range and injectivity)
is a subspace of and is a subspace of . Moreover is injective if and only if . If , then is the null space of and is the column space of .
Proof
(Closure under combinations, and injectivity by differences)
, and if then , so the kernel is a subspace. For the range, lies in it, and shows closure. If is injective and , then . Conversely, if and , then , so . Finally is the equation , and ranges over all combinations of the columns of .
Matrices with respect to other bases
The standard matrix uses the standard basis. For general finite-dimensional spaces, let be linear, let be an ordered basis of and an ordered basis of . Let denote the coordinate column of . The matrix has -th column , and the same argument as for the standard matrix gives
Different choices of bases give different matrices for the same map; relating them is the subject of change of basis (6.8) and, later, of similarity of matrices.
Worked examples
Worked example 6.9.7
(Building a matrix from data)
A linear map satisfies and . Find .
Write and . By linearity,
So . Check: and .
Worked example 6.9.8
(Composition, and order matters)
Let be the rotation of the plane through counterclockwise, with (since and ), and let be the reflection in the line , with . Then
Thus "rotate, then reflect" is the reflection in the -axis: indeed . "Reflect, then rotate" is the reflection in the -axis: . The two composites are different linear maps.
Worked example 6.9.9
(Kernel and range from a matrix)
Let and . The equations , give after subtraction, so and . Hence , which is nonzero, so is not injective. The first two columns and are linearly independent, so and is surjective.
Common mistakes and subtle points
Common mistake
(Reading a composite from left to right)
In the right factor acts first. With and from the example above, is "rotate, then reflect" and has , whereas is "reflect, then rotate" and sends . Reading as " first" gives the wrong map.
Common mistake
(Writing the images as rows)
The columns of are the images of the basis vectors, not the rows. For we have and , so . Placing the images as rows gives , which sends to instead of .
Common mistake
(Swapping domain and codomain in the rank test)
For an matrix the domain is and the codomain is . The map is surjective when the rank is and injective when the rank is . A matrix has rank at most , so it can never be injective, however it is chosen, although it may be surjective.
Affine maps with are not linear, as they fail ; and both additivity and homogeneity are needed, or equivalently the combined condition .
Exercises
Exercise 6.9.10
(Show that is linear, find , and decide whether is injective.)
Core
Solution(3)
- 1
Check linearity
Each coordinate of is a linear expression in with no constant term, so (equivalently, for the matrix found next).
- 2
Read off the columns
Key step
, , , so
- 3
Solve
The second row gives , then the third gives , and the first gives . So and is injective (and, being a map with independent columns, also surjective).
Exercise 6.9.11
(Find the standard matrix of the orthogonal projection of onto the line , given by , and verify that .)
Foundation
Solution(2)
- 1
Compute the images of the basis vectors
and , so .
- 2
Square the matrix
Every entry of equals , so .
Exercise 6.9.12
(Find the matrix of the map reflect in the -axis, then rotate through counterclockwise, and describe the map geometrically.)
Core
Solution(3)
- 1
Write the two matrices
The reflection in the -axis has matrix and the rotation has matrix .
- 2
Multiply, with the rotation on the left
Key step
Because: The reflection acts first, so its matrix is the right factor.
- 3
Identify the map
This is , the reflection in the line . Check on : reflect to , rotate to , which agrees.
Exercise 6.9.13
(Let be linear. Show that if are linearly dependent in , then are linearly dependent in . Give an example showing that the converse fails.)
CoreProof
Solution(3)
- 1
Start from a dependence relation
There are scalars , not all zero, with .
- 2
Apply
, with the same coefficients, not all zero. Hence the images are linearly dependent.
Because: Linear maps preserve combinations and send to .
- 3
The converse fails
∎The zero map sends the independent list in to the dependent list .
Exercise 6.9.14
(Show that is invertible and find .)
Foundation
Solution(2)
- 1
Write the matrix and test invertibility
has determinant , so it is invertible, and is bijective.
- 2
Invert
Check: . Hence .