The limit point and two tangent vectors of a ring's vertex, from the
dominant left eigenvectors of the ring's own subdivision matrix.
Refining a one-ring is linear in its points, so it is a square matrix S
over the centre, edge neighbours and diagonals, built here by refining one
unit ring per point with stepSubdivisionRing. Because the rules are affine,
S keeps constants, so its dominant eigenvalue is 1 and the matching left
eigenvector, scaled to sum 1, is the limit-position mask; the two dominant
left eigenvectors among the masks that sum to zero span the tangent masks
(Halstead, Kass and DeRose 1993 derive the smooth case in closed form). Both
are found by iteration on S from the left: plain power iteration for the
limit, and orthogonal iteration of a two-dimensional zero-sum subspace for
the tangents, re-orthonormalised every step so the weaker tangent is never
lost to the stronger one. limitSubdivisionRing uses this for darts and
for crease sides of three faces, whose sharp edges give them other matrices
than the published masks assume. The tangent search converges at the ratio
of the third to the second dominant zero-sum eigenvalue, about 0.94 for a
dart of valence 12, so EIGEN_STEPS iterations reach rounding.
The limit point and two tangent vectors of a ring's vertex, from the dominant left eigenvectors of the ring's own subdivision matrix.
Refining a one-ring is linear in its points, so it is a square matrix
Sover the centre, edge neighbours and diagonals, built here by refining one unit ring per point withstepSubdivisionRing. Because the rules are affine,Skeeps constants, so its dominant eigenvalue is 1 and the matching left eigenvector, scaled to sum 1, is the limit-position mask; the two dominant left eigenvectors among the masks that sum to zero span the tangent masks (Halstead, Kass and DeRose 1993 derive the smooth case in closed form). Both are found by iteration onSfrom the left: plain power iteration for the limit, and orthogonal iteration of a two-dimensional zero-sum subspace for the tangents, re-orthonormalised every step so the weaker tangent is never lost to the stronger one.limitSubdivisionRinguses this for darts and for crease sides of three faces, whose sharp edges give them other matrices than the published masks assume. The tangent search converges at the ratio of the third to the second dominant zero-sum eigenvalue, about 0.94 for a dart of valence 12, soEIGEN_STEPSiterations reach rounding.