The control polygon

The handles are scaffolding, not structure

Pull a handle in any vector drawing tool and the curve bends away from your finger. The handle's endpoint is not on the curve — it never touches it — and that gap is not a design oversight. It is the geometry working as intended.

The structure doing the work is called the control polygon: the sequence of straight line segments connecting a curve's control points. For a cubic Bézier — the most common case — there are four control points, three segments, and one curve that wrings itself into shape underneath. Pierre Bézier formalised this construction for Renault's body-design workflow in the 1960s; his paper gave designers a way to describe a smooth surface with a small set of movable anchors, published from Boulogne-Billancourt in 1972. The control polygon was always part of the picture, visible on his drawing board if not yet on a screen.

What the polygon guarantees is a family of constraints that hold without any extra code. The curve always starts at the first control point and ends at the last. It never strays outside the convex hull — the tightest polygon that encloses all four points — which means the curve is bounded in a way that is easy to reason about. The tangent at the start runs exactly along the first segment of the polygon; the tangent at the end runs along the last. These are not approximations. They follow directly from the Bernstein polynomial basis that defines the curve, and they are why moving a single handle produces a predictable, local rotation of the curve's shoulder rather than a global distortion.

Subdivision and the polygon that closes in

The deeper power of the control polygon appears when you subdivide. Paul de Casteljau, working independently at Citroën in the late 1950s, discovered that a Bézier curve can be split at any parameter value by repeatedly taking midpoints — or any weighted interpolations — along the polygon's segments. Each split produces two new control polygons, one for each half of the curve. Repeat the process and the polygons converge to the curve itself: the control polygon is the curve's limit, tightening around it like a net. This is why subdivision is both the computational heart of curve evaluation and the basis of adaptive rendering — the renderer keeps splitting until the polygon's segments are flat enough to draw as straight lines without visible error.

The convex-hull property also makes intersection tests cheap. If the convex hulls of two control polygons do not overlap, the curves cannot cross. Subdivide each polygon, test again, and the intersection narrows to a tiny region in a logarithmic number of steps. A great deal of the efficiency in path-based rendering — clipping, hit-testing, boolean operations — traces back to this one geometric fact about the polygon, not the curve.

Why the interface feels the way it does

When Ivan Sutherland built Sketchpad at MIT Lincoln Laboratory in 1963, constraints between drawn objects were already the central idea; the notion that geometry could be steered by a small number of editable parameters rather than redrawn from scratch was already in the air. The control polygon is that idea made concrete for curves. The handles you drag are the polygon's interior vertices — the two middle control points of a cubic — offset so they are visually distinct from the anchor points that sit on the curve.

The asymmetry between anchors and handles, between points on the curve and points that only pull it, is what makes the interface legible. You always know which things are fixed and which are levers. The convex-hull containment guarantee means a pulled handle can never send the curve somewhere unexpected; the curve stays inside the cage you built. That predictability is not accidental. It is what the control polygon was always for.

Pull a handle in any vector drawing tool and the curve bends away from your finger.