Why splines beyond a single Bézier curve
A single cubic Bézier curve has four control points. To describe a long path one joins several Bézier segments, and the previous unit showed that smoothness at each join then becomes a condition on the control points that the designer must enforce by hand. A B-spline curve instead builds the whole path from one list of control points and a single sequence of blending functions that have the required smoothness built in. Each blending function is nonzero on only a few consecutive parameter intervals, so moving one control point changes the curve only locally. This unit uses the construction of Bézier curves, the cubic Bernstein weights and the continuity conditions at joins from the unit on Bézier curves.
Definitions
Definition 5.2.1
(Knot vector and B-spline basis functions)
Fix a degree . A knot vector is a nondecreasing sequence . The B-spline basis functions of degree are defined by the Cox–de Boor recursion:
with the convention that a term whose denominator is zero is taken to be .
Definition 5.2.2
(B-spline curve, domain and knot multiplicity)
For control points the knot vector must have , and the B-spline curve is
Define the domain as . A knot of multiplicity is a value that appears times in the knot vector. If the knots are equally spaced the curve is called uniform.
Basic properties
Theorem 5.2.3
(Local support, nonnegativity and partition of unity)
For every degree and every :
- (Local support) for .
- (Nonnegativity) for all .
- (Partition of unity) if with , then , and all other vanish at this . Hence on the domain.
Proof
(Induction on the degree with a regrouped recursion)
(1) and (2) follow by induction on : the case is immediate, and the recursion expresses as a combination, with nonnegative coefficients on the relevant range ( for the first term, for the second), of and , whose supports lie in and .
(3) Fix . By (1) only the functions with can be nonzero, that is . We show by induction on that these sum to . For the sum is . For write , so that the second coefficient in the recursion for equals (both have denominator and numerators adding to it). Then
Now is supported in and vanishes at , while is supported in and vanishes at . What remains is by the induction hypothesis (these are the functions of degree that can be nonzero on this interval). Terms with zero denominators multiply functions that are identically zero at , so the convention is harmless.
Proof X-ray
Why the coefficients regroup to one
The non-routine step is the regrouping. Each degree- function takes a fraction of one degree- function and a fraction of the next, so in the sum over each interior receives the total coefficient . The proof is a telescoping of coefficients, not a computation with the basis functions themselves, which is why it works for arbitrary knots.
Two hypotheses enter. The condition is used twice: through local support it reduces the sum to terms, and it kills exactly the two leftover boundary terms, because vanishes at and vanishes at . The condition guarantees that all functions belong to the list, so none is missing from the sum. The induction keeps the same interval index while the degree drops, so the hypothesis is applied to the same .
The proof closes at degree , where exactly one indicator function equals at . If or , some of the functions are not in the list and the cancellation leaves a deficit, as the counterexample below shows.
Consequences. On each interval in the domain the curve is a convex combination of the control points , so it lies in their convex hull; and moving changes only for (local control). Without proof we record the smoothness: on a knot of multiplicity , the curve is . A B-spline curve of degree with simple knots is thus , whereas Bézier segments joined at arbitrary controls are only .
Counterexample mode
The partition of unity fails outside the domain
A plausible claim is that the basis functions sum to at every parameter in the knot range. Take , and the knot vector , so the domain is . At only can be nonzero, since and are supported in and . The recursion gives and , so . The sum is . The identity holds exactly on the domain , where all functions that are nonzero on the interval exist. Outside it some of those functions would need an index below or above , so they are missing from the sum.
The uniform cubic B-spline
Take and integer knots. On a unit interval, write with and let be the four relevant control points (). The four basis functions are
and the segment is . These satisfy (the numerators add to ) and they are nonnegative on . Differentiating gives the endpoint values:
Theorem 5.2.4
(Uniform cubic B-splines join with continuity)
The uniform cubic B-spline is at every interior knot: , and .
Proof
(Matching the endpoint formulas of consecutive segments)
Replacing by in the formulas for , and gives exactly , and , which are the values of , , at .
Conversion to Bézier form. Matching endpoints and derivatives (, ), the segment is the cubic Bézier curve with controls
This lets Bézier evaluation and subdivision code draw a uniform cubic B-spline.
Worked examples
Concept lensAnalytic
The cubic basis as repeated averaging
The shape of the uniform cubic basis function is the four-fold convolution of the indicator of with itself. Each convolution adds one order of smoothness, which is why the cubic piece is . Its values at spacing across a knot interval are the weights above: at of the first piece, of the second, and at the center knot, matching , and .
Worked example 5.2.5
(One uniform cubic segment)
Let , , , and take the segment with . At the weights are , , , , which sum to . Hence
The endpoints are and ; the curve does not pass through any control point. The Bézier controls are , , , , and the Bézier midpoint agrees.
Worked example 5.2.6
(A join between two segments)
Take controls , , , , . The join between the first segment (using ) and the second (using ) has position , derivative and second derivative , evaluated from either side; this is the property of the uniform cubic join theorem.
Worked example 5.2.7
(Cox–de Boor on a clamped quadratic)
Let and knots , so and (five control points) and the domain is . Evaluate the basis at , where the nonzero functions are . Start with . Then
and then
They sum to . With , , the curve point is .
Common mistakes and subtle points
Common mistake
(Expecting the curve to pass through its control points)
A B-spline curve does not interpolate its control points in general. The uniform cubic passes through , a weighted average. To start at the first control point, clamp the knot vector by repeating the end knots times, as in the clamped quadratic example.
Common mistake
(Using the whole knot range as the domain)
The domain is , not the whole knot range. Outside it fewer than basis functions are nonzero and the partition of unity fails. The half-open convention means the last parameter value is attained as a limit.
Common mistake
(Reading parameter smoothness as geometric smoothness)
in the parameter is not geometric smoothness at a point where ; a cusp can occur even in a curve.
Three further points are worth keeping in view.
- Raising the multiplicity of a knot lowers smoothness: multiplicity gives a (corner) point and multiplicity allows a break. It also makes some denominators in the recursion vanish, hence the convention.
- Local control cuts both ways: a control point affects only intervals, but a change of one point still changes the curve everywhere in those intervals, not just at one place.
- The knot spacing is part of the curve. Scaling all knots by a factor does not change the geometric curve, but it changes the parameter speed and the derivative vectors.
Exercises
Quick check
(For the uniform cubic segment with , , , , find , , and .)
Use the endpoint formulas with .
Solution · Answer
and . and .
Quick check
(Compute the four uniform cubic weights at and verify that they sum to .)
Use the common denominator .
Solution · Answer
, , , . The numerators satisfy , so the sum is .
Quick check
(A uniform cubic B-spline has many control points. How many segments change when one interior control point is moved?)
Count the segments whose four control points include the moved one.
Solution · Answer
Control point appears in the segments built from with , which are four segments, provided is far enough from the ends that all four exist.
Quick check
(Find the Bézier control points of the uniform cubic segment with , , , .)
Use the conversion formulas with .
Solution · Answer
; ; ; .
Quick check
(For the clamped quadratic knot vector , evaluate the nonzero basis functions at .)
Locate in first, then start from .
Solution · Answer
The interval is with . Then , . Then , , and (the second term has a zero denominator and is ). The sum is .