-
(i) McLachlan's first method (McLachlan, 1972, 1982) is iterative and conceptually the simplest. It sets in which A and R are both orthogonal with R being a current estimate of D and A being an adjustment which, at the beginning of each cycle, has a zero angle associated with it. One iterative cycle estimates a non-trivial A, after which the product AR replaces R. and therefore For this to vanish for all possible rotation axes l the vector must vanish, i.e. at the end of the iteration R must be such that the matrix is symmetrical. The vector g represents the couple exerted on the rotating body by forces acting at the atoms. Choosing gives the greatest and vanishes when in which N is constructed from the current R matrix. A is then constructed from l and this θ and AR replaces R. The process is iterative because a couple about some new axis can appear when rotation about g eliminates the couple about g.
Note that for each rotation axis l there are two values of θ, differing by π, which reduce to zero, corresponding to maximum and minimum values of E. The minimum is that which makes positive. Adding π to θ alters R and N and negates this quantity.
Note, too, that the process is essentially characterized as that which makes the product RM symmetrical with R orthogonal. We return to this point in (iii).
-
(ii) Kabsch's method (Kabsch, 1976, 1978) minimizes E with respect to the nine elements of D, subject to the six constraints by using an auxiliary function in which L is symmetric containing six Lagrange multipliers. The Lagrangian function then has minima with respect to the elements of D at locations which are dependent, inter alia, on the elements of L. By suitably choosing L a minimum of G may be brought into coincidence with the constrained minimum of E. A minimum of G occurs where and the matrix is positive definite, block diagonal, and has which is symmetrical. Thus L must be chosen so as to make the symmetric matrix such that with D orthogonal, or with R replacing D since we are now confined to the orthogonal case, and N is symmetric and positive definite.
-
(iii) Comparison of the Kabsch and McLachlan methods. Using the initials of these authors as subscripts, we have seen that the Kabsch solution involves solving for an orthogonal matrix given that is symmetrical and positive definite. Thus and
By comparison, the McLachlan treatment leads to an orthogonal R matrix satisfying in which is also symmetric and positive definite, which similarly leads to
These seemingly different expressions for and are, in fact, equal, as the following shows therefore Multiplying on both sides by gives and since both N matrices are positive definite and conversely therefore
-
(iv) Diamond's first method. This method (Diamond, 1976a) differs from the previous ones in that the transformation D is allowed to be a general transformation which is then factorized into the product of an orthogonal matrix R and a symmetrical matrix T. The transformation of x to fit X is thus interpreted as the combination of homogeneous strain and pure rotation in which x is subjected to strain and the result is rotated. Furthermore, the solution for D is (in the notation of Kabsch), so that which may be compared with the results of the previous paragraph.
Although this R matrix by itself (i.e. applied without T) does not produce the best rotational superposition (i.e. smallest E), it is the one which exactly superposes the only three vectors in x whose mutual dispositions are conserved, on their equivalents in X, so that the rotation so found is arguably the best defined one.
Alternatives based on , , are all easily developed, and these ideas are developed by Diamond (1976a) to include non-homogeneous strains also.
-
(v) McLachlan's second method. This method (McLachlan, 1979) is based on the properties of the matrix and is immune to singularity of M. If p and q are three-dimensional vectors such that is an eigenvector of this matrix then
If q is negated the second equality is maintained provided λ is also negated. Therefore an orthogonal matrix (consisting of partitions) exists for which in which is diagonal and contains non-negative eigenvalues. The reverse transformation shows that and multiplying the eigenvectors together gives Therefore but is orthogonal and is symmetrical, therefore [by paragraphs (i) and (iii) above] is the required rotation. Similarly, forming corresponds to the Kabsch formulation [paragraphs (ii) and (iii)] since is symmetrical and the same rotation, , appears.
Note that the determinant of the orthogonal matrix so found is twice the product of the determinants of H and of K, and since the positive eigenvalues are collected into it follows that the sign of the determinant of M is the same as the sign of the determinant of the resulting orthogonal matrix. If this is negative it means that the best superposition is obtained if one vector set is inverted and that x and X are of opposite chirality.
Expanding the expression for E, the weighted sum of squares of errors, for an orthogonal transformation shows that this is least when the trace of the product RM is greatest. In this treatment Hence, if the eigenvalues in and − are arranged in decreasing order of modulus, and if the determinant of M is negative, then exchanging the third and sixth columns of produces a product with positive determinant (i.e. a proper rotation) at minimum cost in residual. Similarly, if M is singular and one or more eigenvalues in vanishes it is necessary only to complete an orthonormal set of eigenvectors such that the determinants of H and K have the same sign.
-
(vi) MacKay's method. MacKay (1984) was the first to consider the rotational superposition problem in terms of the vector r of Section 3.3.1.2.1. Using quaternion algebra he showed that if a vector x is rotated to then where and the direction of r is the axis of rotation, as may also be shown from elementary considerations. MacKay then solves this for the vector r by least squares given the vector pairs X and x. The individual errors are and Setting gives which reduces to in which Thus a direct solution for r is obtained, the elements of which are u, v and w, and may be used to construct the orthogonal matrix as in Section 3.3.1.2.1. may be obtained directly from .
If the requisite rotation is 180°, is singular and cannot be inverted. In this case any row or column of the adjoint of is a vector in the direction of the axis. Normalizing this vector to unity, giving l, gives the requisite orthogonal matrix as
Note that MacKay's residual E is quadratic in r. E therefore has one minimum and no maximum, and the minimum is reached on the first cycle of least squares. By contrast, the objective function E that is minimized in methods (i), (ii), (v) and (vii) has one minimum, one maximum and two saddle points in the space of the vector r, as shown in (vii).
It may be shown (Diamond, 1989) that if MacKay's solution vector r is denoted by and that given by the other methods [except (iv)] by then in which A and B are real symmetric, positive semi-definite. A is positive definite unless all the individual vector sums are parallel, as can happen when the best rotation is 180°. Thus the MacKay method only gives the same result as the other methods if:
-
(vii) Diamond's second method. This is closely related to MacKay's method, but uses a four-dimensional vector with components λ, μ, ν and σ in which λ, μ and ν are the direction cosines of the rotation axis multiplied by and σ is . In terms of such a vector Diamond (1988) showed that in which E is the weighted sum of squares of coordinate differences, as before, is its value before any rotation is applied and P is the matrix The rotation matrix R corresponding to the vector is then the last of the forms for R given in Section 3.3.1.2.1.
The minimum E is therefore minus twice the largest eigenvalue of P since , and a stationary value of E occurs when is any of the four eigenvectors of P. E thus has a maximum, a minimum and two saddle points, in general, and its value may be determined before any coordinates are transformed. Diamond also showed that the orientations giving these stationary values are related by the operations of 222 symmetry. Equivalent results have also been obtained by Kearsley (1989).
As an alternative to solving a eigenproblem, Diamond also showed that the vector r, as in MacKay's solution, may be obtained by iterating which has the property that if X and x are exactly superposable then is the exact solution and no iteration is necessary. If X and x are similar but not exactly superposable then a small number of iterations may be required to reach a stable r vector, though the matrix is not required. As in MacKay's solution, is singular at the end of the iteration if the required rotation is 180°, but the MacKay and Diamond methods both have the advantage that improper rotations are never generated by these means, and methods based on P and rather than Q and r are trouble-free for 180° rotations. The iterative loop in this method does not require Rx to be redetermined on each cycle.
Finally, it may be shown that if are the eigenvalues of P arranged in descending order and is negative, then a closer superposition may be obtained by reversing the chirality of one of the vector sets, and the R matrix constructed from optimally superimposes Rx onto − X, the enantiomer of X (Diamond, 1990b).
|