Motivation
A graphics path should be easy to edit while remaining mathematically predictable. Moving a control point should have a clear effect on the curve; joining two pieces should not unexpectedly introduce a corner. Cubic Bézier curves address these needs using repeated affine interpolation. They complement the existing chapters on transforms, rasterization, lighting and ray visibility: here the object being represented is a curve, not a pixel or a visibility test.
From control points to a curve
Definition 5.1.1
(Cubic Bézier curve)
For control points and , define
The four scalar coefficients are the cubic Bernstein weights. The endpoints are and ; the middle control points need not lie on the curve.
The weights are nonnegative and sum to . Hence each curve point is a convex combination of the four controls, so the entire segment lies inside their convex hull. The result holds for ; extrapolating outside that interval can make weights negative. A convex-hull enclosure does not by itself guarantee that the curve is one-to-one or has no repeated points.
Common mistake
(Reading control points as points on the curve)
Only and lie on the curve. In the cubic with controls used below, the middle controls have height , but the curve's height is , so the curve never reaches them. The controls shape the curve and enclose it; they do not lie on it.
Repeated interpolation and subdivision
Write . Define first-level points for , then for , and finally . Every step is an affine combination, so applying an affine transformation to the controls before construction gives the same result as transforming the finished point.
Theorem 5.1.2
(The repeated interpolation computes the Bézier point)
With , and constructed from the controls at parameter as above, .
Proof
(Expanding the three levels)
By definition . Then
Finally , whose coefficients are for ; for ; for ; and for . These are the Bernstein weights of the definition.
Worked example 5.1.3
(Halfway through a symmetric cubic)
Take , , and . At , the first level is , , . The second is , , and . By the subdivision theorem below, these values also supply both subdivided control polygons. Although the middle controls have height , the midpoint of this curve has height .
Concept lensAlgebraic
The same curve in the power basis
Expanding the Bernstein weights gives
The higher coefficients are multiples of the first, second and third differences of the controls. For the worked cubic this is , which gives again. The power form is convenient for evaluation, while the Bernstein form keeps the control polygon and convex-hull enclosure visible.
Derivatives and joining segments
Differentiating the cubic gives
In particular and . Comparing with the expressions for and above shows that the bracket equals , so . For the example, the endpoint derivatives are and ; at the derivative is , equal to . A zero derivative does not provide a tangent direction through this derivative vector, although a geometric tangent can still exist. For example, for has , but its normalized forward chord direction is for every . A derivative-based direction test must therefore check for a nonzero derivative.
Theorem 5.1.4
(Subdivision at a parameter)
Fix and construct at . Then the cubic Bézier curve with controls is for , and the curve with controls is .
Proof
(Four Hermite conditions determine the cubic)
Let be the Bézier curve with controls and . Both are polynomials of degree at most in , and a cubic is determined by its values and first derivatives at and . We compare these four vectors.
Values: and , the latter by the theorem on repeated interpolation.
Derivatives: , since . Also , since and .
Hence . The second statement follows by applying the first to the reversed curve with controls at parameter , which has the same construction points in reverse order.
Proof X-ray
Why four conditions suffice and where the construction enters
The proof avoids expanding coefficient by coefficient. Its key move is to count: a cubic has four coefficients, and matching two values and two first derivatives fixes all four, so no further check is needed. Everything else is reading off the four vectors from the construction.
Each of the four comparisons uses one fact from the lerp tree. The endpoint uses the preceding theorem. The two inner controls are placed by the same ratio along the tree: sits a fraction along the first leg , and a fraction along , which is how the factor in and arises. The identity is what ties the right end of the left piece to the original curve.
The argument closes by uniqueness. It would fail for a quartic or for fewer matched data. At the extreme values and one piece degenerates to a single point, and the identity still holds.
Worked example 5.1.5
(Splitting the example cubic at one quarter)
For the cubic with controls take . Then , , . Next and , and . The left piece has controls and the right piece .
Check: has and , which agrees with . The left piece starts with derivative , as the reparameterization requires.
Theorem 5.1.6
(Position and derivative continuity at a join)
Let a first cubic have controls and a second have controls . If each is parameterized on its own unit interval and the intervals join in order, the composite is position-continuous exactly when . It is also at that join exactly when . For nonzero endpoint derivatives, requiring only a common oriented tangent direction is weaker: one may be a positive scalar multiple of the other.
Proof
(Comparing one-sided positions and derivatives)
The first segment ends at with left derivative ; the second starts at with right derivative . The composite is continuous at the join exactly when the one-sided positions agree, , and it is there exactly when, in addition, the one-sided derivatives agree, . If the derivatives are nonzero and one is a positive multiple of the other, they point in the same direction, which is the weaker condition.
For the example above, take and ; then both one-sided derivatives are . The remaining controls determine the rest of the second segment. Equal directions alone can hide a speed jump. If the first and second local unit parameters are instead mapped to time intervals of lengths , the time derivatives are and ; compare these scaled vectors for time-based continuity.
Common mistake
(Matching directions but not speeds)
Take the first segment of the example, with , and a second segment with and . Then points in the same direction but has half the length. The path has a common tangent line but the velocity jumps from to , so it is not in the parameter.
Worked example 5.1.7
(A join between segments of unequal duration)
Suppose the first segment of the example runs over time length and the second over . The first has end velocity . For a time-based join with we need , so and . Using the unit-interval rule instead would give velocity , a jump. Parameterization, not only control geometry, defines the continuity.
Counterexample mode
A shared endpoint need not make a smooth join
Set , and . The segments meet, so position is continuous, but the incoming derivative is and the outgoing derivative is . The join has a corner. Sharing an endpoint alone is not a derivative-continuity test.
Rendering limits
Subdivision supports curve drawing and bounding because each piece lies in its own control hull. A renderer may recursively subdivide until an explicitly chosen screen-space flatness criterion is met, but any finite tolerance introduces approximation error and must be interpreted in the chosen coordinate system. Affine transformations preserve the construction; after perspective division the projected polynomial curve is generally rational, so applying ordinary affine Bézier interpolation directly to projected controls is not generally exact. Clipping, adaptive sampling and pixel coverage remain separate implementation questions.
Exercises
Quick check
(For the worked cubic with controls , find , , and .)
Use the endpoint formulas.
Solution · Answer
The positions are and ; the derivatives are and .
Quick check
(For controls , find the de Casteljau point at .)
Take midpoints three times.
Solution · Answer
The first level is ; the second is , so the curve point is .
Quick check
(A first cubic ends at with . A second starts at . Find for continuity when both local intervals have unit length.)
Match the one-sided derivatives.
Solution · Answer
Set , giving . Other control points remain free, and unequal durations require the scaled condition.
Quick check
(A point lies inside the convex hull of a curve's controls. Must it lie on the curve? Explain.)
Test the point for the worked cubic.
Solution · Answer
No. For the worked cubic, the control point lies in the hull, but for the curve's height is ; it never reaches height .
Further study
Rational curves, spline bases, higher-order joins, robust flattening and surface patches need further analysis. The results here do not supply a production renderer or prove any particular pixel-coverage rule.