Inside the Matrix

Vectors

A vector is simply a tuple of numbers, either written in a row as

v = (v1, v2, ..., vn),

or in a column as

    /  v1 \
    |  v2 |
v = | ... |.
    \  vn /

For brevity, I'll mostly stick to the row format in this text, but they are absolutely equivalent.

In computer graphics, we can use vectors to hold directions and distances in euclidean space. A vector then contains one number for each dimension of space, representing the distance along a single axis. Thus v = (x, y, z) is only a short form of writing v = x*a_x + y*a_y + z*a_z, where a_x, a_y, and a_z are the three axes of space.

We can also use vectors to represent positions in space, by using a reference point O and storing the distance between a point and the reference point. However, distances and points now look the same, which is a Bad Thing, because they aren't. This is where homogeneous coordinates come into play. We can solve this ambiguity by adding another element, the homogenous coordinate, to the vector. For distances, this coordinate is 0, for points it is 1. (Actually, for points it must only be non-0, but I won't go into all the details of homogeneous coordinates. Suffice to say that a vector v = (x, y, z, w) with w != 0 represents the same point as v' = (x/w, y/w, z/w, 1).)

With this addition, the vector v = (x, y, z, w) is a short form of writing v = x*a_x + y*a_y + z*a_z + w*O. Vectors are always relative to a /coordinate system/, defined by the three axes a_x, a_y, and a_z, and the reference point O, also called the /origin/.

Transformations

Affine transformations, such as scaling, rotation, translation, shearings, and any combination thereof, can be expressed by defining a new coordinate system. Scalings are simply changes to the length of the axes, rotations change the orientation of the axes, translations move the reference point, and shearing change the angles between the axes.

For example, an object in a coordinate system whose axes all have a length of two units, is twice as big as the same object in a coordinate system whose axes all have a length of one unit.

Now, if we express the axes and origin of the transformed coordinate system in terms of the original coordinate system, we can apply the transformation to any vector (point or distance) by simply evaluating the equation given above.

We'll call the transformed coordinate system T, with axes T_x, T_y, and T_z, and origin T_O, and the original coordinate system O, with axes O_x, O_y, and O_z, and origin O_O. The transformed coordinate system T is defined relative to the original coordinate system O:

T_x = (T_xx, T_xy, T_xz, 0)
T_y = (T_yx, T_yy, T_yz, 0)
T_z = (T_zx, T_zy, T_zz, 0)
T_O = (T_Ox, T_Oy, T_Oz, 1)

(Notice how the axes all have a homogeneous coordinate of 0, as they are distances, and the origin has a homogeneous coordinate of 1, as it is a point.)

With these definitions, we get

v' = (x', y', z', w') = x*T_x + y*T_y + z*T_z + w*T_O

   =   x * (T_xx * O_x + T_xy * O_y + T_xz * O_z + 0 * O_O)
     + y * (T_yx * O_x + T_yy * O_y + T_yz * O_z + 0 * O_O)
     + z * (T_zx * O_x + T_zy * O_y + T_zz * O_z + 0 * O_O)
     + w * (T_Ox * O_x + T_Oy * O_y + T_Oz * O_z + 1 * O_O)

   =   (x*T_xx + y*T_yx + z*T_zx + w*T_Ox) * O_x
     + (x*T_xy + y*T_yy + z*T_zy + w*T_Oy) * O_y
     + (x*T_xz + y*T_yz + z*T_zz + w*T_Oz) * O_z
     + (x*   0 + y*   0 + z*   0 + w*   1) * O_O.

Now, this looks rather complicated at first sight, but if you look closely, you'll notice a certain pattern.

Enter the Matrix

People familiar with linear algebra will probably recognize the pattern. It look suspiciously like the product of matrizes, and indeed, it can be written as such.

We have to combine the axes and origin of the transformed coordinate system into a 4x4 matrix M. This can be done in more than one way, but I'll stick to the convention used by OpenGL. The other possibilities lead to different multiplication orders (remember matrix multiplication is not commutative) and vector notations.

If we combine the axes and origin such that each vector occupies one column of the final matrix, and consider the vector as a column matrix (a matrix with only a single column, much like the column notation of vectors), the transformation can be expressed as a simple

v' = M * v

     / T_xx T_yx T_zx T_Ox \   / x \
     | T_xy T_yy T_zy T_Oy |   | y |
   = | T_xz T_yz T_zz T_Oz | * | z |.
     \    0    0    0    1 /   \ w /

Working with Transformations

Now, that we have shown how affine transformations can be expressed as matrix multiplications, it is trivial to show how to combine transformations.

Suppose we want to move and scale an object. We have the scaling transformation stored in the matrix S, and the translation in the matrix T.

v' = T * (S * v)
   = (T * S) * v
   = TS * v

Thus, we can combine both transformations into a single matrix, simply by multiplying the two transformation matrizes. Note, however, that the order is important. T * S moves the scaled object, whereas S * T scales the already moved object, /thus amplifying the movement/.

When using OpenGL, transformations are accumulated from left to right. Thus, to produce the transform T * S, one has to call the gl-functions in the order

glTranslate(...);
glScale(...);

Or, put another way, the first transformation affecting an object, is the transformation last executed.