Use the standard partition normalization of the B-splines, so and the subpartition of unity for B-splines gives . Write the interpolating spline as and the sampled values as . The interpolation equations are , hence whenever the stated interpolation map is well defined. For every ,
The sampling map also satisfies . Taking supremum norms and then the operator norm therefore proves
This is the B-spline interpolation operator norm bound, and it applies on any interval containing the sampling sites on which these normalized B-splines are used.
There is a small qualification to the printed hypotheses: diagonal positivity alone does not ensure uniqueness unless the usual site assumptions are supplied. For example, with two quadratic B-splines on knots and , both diagonal values equal , but has two identical rows and is singular. The intended increasing sites with satisfy the Schoenberg–Whitney theorem, making invertible; alternatively invertibility is implicit in the premise that is the unique interpolation map. The displayed bound is proved under that premise, rather than inferring invertibility from diagonal positivity alone.
For unit-spaced knots, the quadratic cardinal B-spline is a translate of
For example, the Cox-de Boor recurrence gives from the two linear hats. Substituting their linear pieces gives the three expressions above. At , only the B-splines with indices have nonzero values, and their values are respectively
Only indices belong to the given basis, so entries outside that range are omitted at the boundary.
The midpoint quadratic spline collocation matrix is
Its strict diagonal dominance margin is at least . The inverse infinity-norm bound from diagonal dominance thus already proves : choosing an index where a vector attains its maximum absolute entry gives , which also proves invertibility.
To evaluate the finite-dimensional norm exactly, set and . If is the matrix with ones on its two adjacent diagonals and zeros elsewhere, then . Since , the Neumann series
has nonnegative entries. Thus has absolute entry values equal to those of . Its absolute row-sum vector is , characterized by
The homogeneous roots are , and the constant particular solution is . Writing and imposing the two boundary values gives
This is largest at the middle row or rows. Therefore, with , the exact norm is
For example, it is for and for , and it increases towards two as the system grows. Thus two is a uniform bound, rather than the exact finite- inverse norm.
Combine the B-spline interpolation operator norm estimate from part 1 with the inverse operator norm computed in part 2(b):
so in particular , uniformly in . The subpartition of unity for B-splines is important at the outer boundary: the finite basis may sum to less than one there, but that only strengthens the estimate.

Articles by others on the same topic (0)

There are currently no matching articles.