Convex rational cubic G2 Hermite interpolation of a short conic end.
Degree-elevating a rational quadratic gives weights [1,u0,u0,1].
For a fixed control polygon, start curvature is proportional to v/u^2
and end curvature to u/v^2. Thus [1,u0/tau^(2/3),u0/tau^(1/3),1]
changes only the start curvature by tau. This is the standard rational
cubic osculatory problem described by Habib and Sakai (2001, p341),
https://www.jams.or.jp/scm/contents/Vol-5-5/5-41.pdf .
The original segment and this positive-weight convex cubic lie between
its chord and first tangent. For T<=1/2 their same-ray radial separation
is at most (a-x(T))/cos(phi(T)) <= 4T^2hypot(a,b). The chosen T is
smaller than the square root of that error budget. It also vanishes
linearly with 1-tau. Near tau=1, the radial correction is
O((1-tau)*T^2), its phase derivative O((1-tau)*T), and its second phase
derivative O(1-tau). Scaling the compact angular interval then makes all
mixed derivatives of total order at most two tend to zero. Away from
that branch the positive regular rational inverse is smooth.
Small x deficits are stored separately from a, avoiding subtraction of
nearly equal world coordinates before differentiating.
Convex rational cubic G2 Hermite interpolation of a short conic end.
Degree-elevating a rational quadratic gives weights [1,u0,u0,1]. For a fixed control polygon, start curvature is proportional to v/u^2 and end curvature to u/v^2. Thus [1,u0/tau^(2/3),u0/tau^(1/3),1] changes only the start curvature by tau. This is the standard rational cubic osculatory problem described by Habib and Sakai (2001, p341), https://www.jams.or.jp/scm/contents/Vol-5-5/5-41.pdf .
The original segment and this positive-weight convex cubic lie between its chord and first tangent. For T<=1/2 their same-ray radial separation is at most (a-x(T))/cos(phi(T)) <= 4T^2hypot(a,b). The chosen T is smaller than the square root of that error budget. It also vanishes linearly with 1-tau. Near tau=1, the radial correction is O((1-tau)*T^2), its phase derivative O((1-tau)*T), and its second phase derivative O(1-tau). Scaling the compact angular interval then makes all mixed derivatives of total order at most two tend to zero. Away from that branch the positive regular rational inverse is smooth.
Small x deficits are stored separately from a, avoiding subtraction of nearly equal world coordinates before differentiating.