Numerically reproduce the endpoint-bitangent equioscillation construction of
Ghosh, Petitjean and Vegter, Approximation by Conic Splines (2007), sections
3.2–3.4 and 6, DOI 10.1007/s11786-007-0004-8.
Source breaks are hard boundaries: piecewise Hermite input does not acquire
the paper's smoothness merely by supplying derivatives. On each smooth chart
we sample nonzero curvature and monotone affine curvature, construct the
endpoint triple-contact limits of the positive-weight bitangent family, and
bisect its weight until refined positive and negative normal extrema balance.
Located extrema are exchanged into the fitting nodes; independent held-out
nodes then validate the selected coefficient rather than driving its fit.
Endpoint bisection then selects an interval meeting the requested tolerance.
Unlike the paper's symbolic implementation, this experiment uses finite
admission/extremum samples, so neither Hausdorff certification nor global
minimum arc count is claimed. Collinear charts are an explicit line extension.
A zero-curvature chart endpoint uses the compactified infinite-weight limit
only to bracket finite positive weights. This is an explicit numerical
boundary extension, not a proof that the paper's theorem admits that endpoint.
Source callback failures refuse the whole call and discard all coefficients;
sampled mathematical admission failures may leave explicitly partial coverage.
Numerically reproduce the endpoint-bitangent equioscillation construction of Ghosh, Petitjean and Vegter, Approximation by Conic Splines (2007), sections 3.2–3.4 and 6, DOI 10.1007/s11786-007-0004-8.
Source breaks are hard boundaries: piecewise Hermite input does not acquire the paper's smoothness merely by supplying derivatives. On each smooth chart we sample nonzero curvature and monotone affine curvature, construct the endpoint triple-contact limits of the positive-weight bitangent family, and bisect its weight until refined positive and negative normal extrema balance. Located extrema are exchanged into the fitting nodes; independent held-out nodes then validate the selected coefficient rather than driving its fit. Endpoint bisection then selects an interval meeting the requested tolerance. Unlike the paper's symbolic implementation, this experiment uses finite admission/extremum samples, so neither Hausdorff certification nor global minimum arc count is claimed. Collinear charts are an explicit line extension. A zero-curvature chart endpoint uses the compactified infinite-weight limit only to bracket finite positive weights. This is an explicit numerical boundary extension, not a proof that the paper's theorem admits that endpoint. Source callback failures refuse the whole call and discard all coefficients; sampled mathematical admission failures may leave explicitly partial coverage.