1 Introduction Today were going to discuss linear transformations. These are the functions between vector spaces that preserve vector addition and scalar multiplication, the two fundamental opera- tions of vector spaces. So far, were just discussing vector spaces of the form R n for some natural number n (euclidean spaces), and so the linear transformations well be studying (for now) will be of the form T : R k R n , but soon we will consider more general vector spaces and linear transformations (starting in chapter 4). Our plan for today is as follows. First, well motivate linear transformations by looking at matrix transformations, which came up while looking at systems of linear equations. Then well dene linear transformations and observe that in fact theres a close relationship between matrices and linear transformations. Well consider geometrically some examples of certain types of linear transformations. Finally well consider the two functional properties of one-to-one and onto and characterize which linear transformations have these properties. 2 Matrix Transformations Recall that a system of linear equations can be expressed as a matrix equation Ax = b, where A is a matrix, b is a column vector, and x is a column vector of variables. A solution to the system is some actual vector x such that Ax = b. Another way to express this is functionally. Suppose that A is an nxk matrix. Then x Ax is a function which takes in a vector x R k as input, and spits out a vector Ax R n as output. This function x Ax is called a matrix transformation. A solution of the system Ax = b is a vector that, when input into this transformation, becomes b. For example, consider the system Ax = b where A =
1 2 1 1 0 1
and b =
1 2
. Then the matrix transformation x Ax is a function with domain R 3 and codomain R 2 , and a solution to the system is just some vector
x 1 x 2 x 3
such that when it is used as input, the
matrix transformation produces
1 2
as output. Heres a little puzzle for you. Let A be some 2x2 matrix. Suppose we knew that A
1 2
5 3
and A
1 1
1 1
. Is it possible for us to gure out A
0 3
? Yes, it is. Note that
1
1 2
1 1
0 3
. Then we can reason as follows:
A
0 3
= A
1 2
1 1
= A
1 2
+ A
1 1
5 3
1 1
6 2
The property were using of matrix multiplication that I want to point out in this reasoning is A(x + y) = Ax + Ay. We summarize this property in words by saying that matrix mul- tiplication distributes over vector addition, or that matrix transformations preserve vector addition. You should be familiar from elementary school how multiplication distributes over addition (e.g., 7 (100 + 9) = 7 100 + 7 9) and this is the same idea. It doesnt matter whether you add rst and then multiply, or multiply and then add. Or expressed in terms of transformations: it doesnt matter whether you add rst and then apply the matrix trans- formation, or apply the matrix transformation and then add. Matrix transformations also satisfy a similar property involving scalar multiplication. Specically A(cx) = c(Ax) for any scalar c. We say that matrix transformations preserve scalar multiplication. For example, A
2 2
= 2A
1 1
. 3 Linear Transformations A linear transformation fromR k to R n is a function T : R k R n such that for every x, y R k and c R we have T(x+y) = T(x) +T(y) and T(cx) = cT(x). In words, T preserves vector addition and scalar multiplication. Matrix transformations x Ax are examples of linear transformations. Are there other examples? Well, yes and no. It turns out that every linear transformation from R k to R n can be represented by an nxk matrix. However, this does not imply that linear transformations and matrix transformations are exactly the same thing. First of all, later we will be interested in linear transformations whose domain and codomain are not euclidean. E.g., the derivative is a linear function from the vector space of (innitely) dierentiable functions to itself. Also, as well see later, more than one matrix can represent the same linear transformation (albeit with respect to a dierent coordinate system). With all that said though, there is a unique standard matrix associated with every linear transformation between euclidean spaces. Suppose that T : R k R n is linear. Is T given by matrix multplication for some matrix A? I.e., is there a matrix A such that T(x) = Ax for every x R k ? Well, if there is such a matrix, then we better, in particular, have T(e i ) = Ae i for each i = 1, . . . , k. (e i denotes the vector (0, . . . , 0, 1, 0, . . . , 0) with 0s everywhere except in the i th spot.) But observe that Ae i is simply the i th column of A. Thus, the desired matrix has to be, if anything works,
T(e 1 ) T(e 2 ) T(e k )
Well, lets check that this works. Let x =
x 1 . . . x k
R k be any vector. Note that x = x 1 e 1 + + x k e k . Repeatedly using the fact that T preserves vector addition and scalar 2 multiplication yields T(x) = T(x 1 e 1 + x 2 e 2 + x k e k ) = x 1 T(e 1 ) + x 2 T(e 2 ) + + x k T(e k ) =
T(e 1 ) T(e 2 ) T(e k )
x 1 . . . x k
= Ax So we see that this choice of A actually does work. So, given any linear transformation T, to nd the standard matrix associated to it, you just need to calculate T at each of the standard basis vectors e i and make the matrix with the results of these calculations as columns. 4 Geometric Examples Lets think geometrically about some linear transformations on the plane, i.e. from R 2 to R 2 . Rotation (by any angle about the origin) is linear because it doesnt matter if you scale rst and then rotate, or rotate rst and then scale, and similarly it doesnt matter if you add rst and then rotate, or rotate rst and then add you get to the same place. What is the standard matrix for rotation by radians? Remember we just need to gure out where the standard basis vectors e i are sent. Well, rotating (1, 0) by puts it at (cos , sin ) recall that cos and sin are dened in exactly this way. Similarly, (0, 1) gets sent to (cos( + /2), sin( + /2)) which can also be written (sin , cos ). Thus the standard matrix is
cos sin sin cos
. For example, if = /2, then the standard matrix is
0 1 1 0
Horizontal and vertical contractions and expansions are also linear. I.e.,
k 0 0 1
scales horizontally by a factor of k, and
1 0 0 k
scales vertically by a factor of k. If k > 1 we have
expansion, if 0 < k < 1 we have contraction. If k = 0, we actually call it a projection. If 1 < k < 0 we have a reection and a contraction, while if k < 1 we have a reection and an expansion. Really the only other kind of linear transformation on the plane (in a sense to be made precise later) is a shear. The standard matrix for a horizontal shear is
1 k 0 1
, and for a vertical shear
1 0 k 1
. These transformations x one of the axes and then slide points in a
direction parallel to the xed axis. The amount any point slides in this way is determined by how far that point is from the xed axis. One thing to keep in mind when visualizing linear transformations on the plane is that parallelograms are always mapped to parallelograms, in the sense that given four points 3 x, x + a, x + b, x + a + b in the form of a parallelogram, the image of these four points T(x), T(x+a), T(x+b), T(x+a+b), i.e. T(x), T(x) +T(a), T(x) +T(b), T(x) +T(a) +T(b), is also in the form of a parallelogram. Sometimes though, as in the case of projection, things get attened out and so the parallelogram becomes degenerate. 5 One-to-one, Onto Recall that a function f : A B is called one-to-one (also: injective) if for all a, a
A, f(a) = f(a
) implies a = a
. Another way to put this is that a = a
implies f(a) = f(a
). In words, distinct elements of the domain are mapped to distinct elements in the codomain. Questions: Is projection onto the x-axis one-to-one? No, because there are points in the domain that are smooshed to become the same point in the codomain. Is rotation by 90 degrees one-to-one? Yes, because every point has a unique point that rotates to it (to nd that point, rotate backwards by 90 degrees). Note that in the rst case more than one point is mapped to 0, and in the second case, only 0 rotates to 0. In general, if T is a linear transformation, then T is one-to-one i 0 is the only vector T maps to 0. Its always the case that T(0) = 0 for any linear transformation our claim is that if 0 is the only thing that works, then T is 1-1 and vice versa. Lets prove it. Suppose that T is 1-1. Then if T(x) = 0, then x must equal 0 lest T not be 1-1. Now for the other direction. Suppose that T(x) = 0 implies x = 0. Let T(v) = T(w). We want to show that v = w. Well, T(v) T(w) = 0 and so, by the linearity of T, T(v w) = 0. Thus v w = 0, and so v = w. We can say a bit more about this characterization in the case of euclidean spaces. Let T : R k R n be a linear transformation. Let A be the standard matrix for T. I.e., T(x) = Ax for every x R k . Then T is 1-1 i T(x) = 0 implies x = 0 i Ax = 0 implies x = 0 i the columns of A are linearly independent i there is a pivot in every column of the reduced echelon form of A (i.e. there is no free variables). Since the columns of A are T(e 1 ), . . . , T(e k ) we could also say that T is 1-1 i {T(e 1 ), . . . , T(e k )} is linearly independent. Recall that a function f : A B is called onto (also: surjective) if for every b B there is some a A such that f(a) = b. I.e., everything in the codomain is obtained as output under some appropriate input. In other words, the range equals the codomain. In the case T is a linear transformation of euclidean spaces, and A is the standard matrix for T, then T is onto i Ax = b is consistent for every b R n i the columns of A span R n i there is a pivot in every row of the reduced echelon form of A. 4
