Rotations and AntiSymmetric Tensors 

In a previous note we observed that a rotation matrix R in three dimensions can be derived from an expression of the form 



and similarly in any other number of dimensions. The first matrix on the right side is simply the identity matrix I, and the second is a antisymmetric matrix A (i.e., a matrix that equals the negative of its transpose). Thus we can write the above relation, for any number of dimensions as 



Solving for R, we get 



There is no ambiguity in writing the right hand side as a ratio, because the matrices (I – A) and (I + A)^{1} commute, as shown by considering the two products (expanding the inverse matrix into the geometric series) 



When the right hand sides are multiplied out, every term is a product of identity matrices and the A matrix, so the order is immaterial, and hence the two products are identical. To verify that, in fact, every matrix R expressible in the form (2) is a rotation, recall that a rotation (i.e., a special orthogonal transformation) is defined by the conditions 



The first of these two conditions is nearly sufficient by itself because, since the determinants of a matrix and its transpose are equal, the first condition implies that R is either +1 or 1. Hence the second condition serves only to single out the positive sign of the determinant. To show that any R given by (2) is a rotation, we rewrite that equation in the form 



Taking the transpose of both sides, we get 



Now, since A is antisymmetric, it equals the negative of its transpose, so this equation can be written as 



Multiplying both side of (3) on the right by the respective sides of this equation, we get 



Multiplying both sides on the right and left by the inverses of IA and I+A respectively, this becomes 



where we have made use of the commutativity (as noted previously) of the factors on the right side. To prove that the determinant of R is +1 rather than 1, it suffices to observe that equation (3) gives the determinant +1 for R if A = 0, and by continuity it must be +1 for any other (realvalued) matrix A. 

By equation (2) we have 



which is known as Cayley’s transformation, and it can be inverted to give 



We proved above that is A is antisymmetric then R is a rotation. Conversely, if R is a rotation, then A is antisymmetric. To prove this we need to show that A = −A^{T}. Negating the transpose of the above expression for A gives 



Since (A^{T})(A^{−1})^{T} = (A^{−1}A)^{T} = I, we know that (A^{T})^{−1} = (A^{−1})^{T}, so we have 



Now, if R is a rotation we have R^{T} = R^{−1}, so we have 



which proves that A is antisymmetric. Consequently, we have a onetoone mapping between rotations and antisymmetric matrices. This applies to all rotations, including infinitesimal rotations, i.e., rotations of the form ρ = I + ε where ε is an infinitesimal antisymmetric matrix, as confirmed by the fact that we can write ρ = (I + ε/2)/(I − ε/2). 

In the previous note we also discussed a condition on rotation matrices that can be written as 



where k is a constant depending on the number of dimensions of the space. We noted that in two dimensions the determinant of the matrix I  R is nonzero except for any rotations other than the identity, so it can be divided out of this equation, and we have k = 1 with the equality 



For higher dimensions (at least for dimensions three and four) the determinant of I – R is identically zero. Hence we can only factor the relation as 



In the case of three dimensions, the solution is given by setting k = 2, which results in the expression in square brackets reducing to 



where the variables a,b,c are as defined for the threedimension rotation given in (1), and 



The other factor on the left side of (4) is 



It is straightforward to verify that the product of this matrix and the matrix in the preceding equation is identically the zero matrix, which is to say, the rows of one are orthogonal to the columns of the other, and vice versa. We already noted that the determinant of I – R vanishes in more than two dimensions, but it’s interesting that the determinant of the second factor in (2) also vanishes, in three dimensions, provided we set k = 2. Letting f(z) denote the function 



we can plot f(z) versus z for real values of z, as shown in the figure below for two arbitrarily chosen rotations in three dimensions. 


In addition to the common repeated root at z = 2, the function f(z) for each threedimensional rotation R also has another root at 



as can be verified by direct substitution into f(z), then factoring out the I + R/4, and noting that 



In four dimensions the general rotation can again be expressed as R = (I  A)(I + A)^{1}, where now the matrices are of dimension four and the antisymmetric matrix A is 



As discussed in the previous note, relation (4) is satisfied with k = 4, but only under the condition that af – be + cd = 0. (See the note on Lorentz Boosts and the Electromagnetic Field for a comment on this condition.) Just as in the threedimensional case, equation (4) is also satisfied by z = I + R/4, in addition to the general solution z = 4. In this case each of these represents a repeated root, as can be seen from the figure below, which shows f(z) for two rotations in fourdimensional space (each satisfying the special condition af – be + cd = 0). 


More generally, if we allow the quantity af – be + cd to be nonzero, we find that f(z) still has two repeated roots, although now there is no common solution with z = 4. Instead, both of the roots vary, as shown in the figure below for two arbitrarily selected rotations. 


If we apply the homogeneous approach taken in the previous note, we can define a fully general fourdimensional rotation matrix R by the relation 



for constants a, b, c, d, e, f, g, and we then define the eight parameter 



We can now extend the 2parameter expression in two dimensions and the 4parameter expression in three dimensions to an 8parameter expression in four dimensions, giving the completely general fourdimensional rotation matrix as 



where 


The eight parameters are constrained by the normalizing condition, and by the relation between h and the other parameters, so there are six degrees of freedom (just as there are three degrees of freedom for rotations in three dimensions, and one degree of freedom for rotations in one dimension). 
