A Practical Guide To Splines

What Actually Happens When You Define a Spline

You place points. The software connects them with a curve. That is the wrong way to think about it, and it will cost you time later. A spline is a piecewise polynomial function. Each segment between control points is a separate polynomial, and the whole thing is stitched together so that derivatives match at the junctions. A cubic B-spline has C2 continuity — position, tangent, and curvature are all smooth across every knot. That matters when you're toolpaths or animating motion, because G2 discontinuities show up as visible bumps or sudden acceleration spikes. I learned this the hard way three years ago on a CAM project. We were generating paths for a 5-axis mill, and the part had a long gentle curve near the edge. The spline looked fine on screen. On the machine, every third pass left a visible marksline. Took me two hours to realize the issue wasn't the cutter or the feed rate — it was a discontinuous fourth derivative in the spline representation. The curvature was smooth, but jerk in the parameter space was causing the controller to hesitate at knot boundaries. The fix was switching from a uniform knot vector to a centripetal one and recomputing the control points. Saved the job. Cost me a morning I'd already budgeted for something else.

A Practical Guide To Splines

This section covers the parts nobody explains in the tutorial videos. The stuff that actually shows up when you're six hours into a project and the curve won't behave. Control points versus fit points. Most people confuse these. A control point pulls the curve toward it but does not necessarily lie on it. A fit point forces the curve through the coordinate you specify. In NURBS modeling, they behave completely differently under manipulation. If you're editing a curve interactively, understand which type your tool is using before you pull a point and wonder why the shape moved wrong. Fit-point curves give you direct shape control but produce wildly varying polynomial degrees. Control-point curves maintain consistent degree but the influence is indirect and non-local — moving one point affects the entire span around it. Knot vectors are where things break. A clamped knot vector places multiple knots at the endpoints, which forces the curve to pass through the first and last control points. This is what you want for most engineering work. An unclamped or open uniform vector gives you something totally different — the curve floats inside the control polygon and never touches the endpoints. I've seen this cause real problems in reverse engineering. Someone scans a part, fits a spline through the scan points, and the resulting surface doesn't align with the actual datum edges because the knots weren't clamped. The scan data was fine. The knot choice was the mistake.

Renormalizing knot vectors. When you add a control point, the software needs to reparameterize. This is called knot insertion. It's exact in theory but floating point drift creeps in after enough operations. I've worked with files that had 40 or 50 successive insertions, and the tolerance stack made the curve deviate from its intended shape by amounts that showed up only at high magnification. The workaround is to rebuild the spline from scratch using the current control points as constraints, then delete the old chain. Takes about 30 seconds and prevents compounding error. C0, G1, G2 — and when each actually matters. C0 means the curve is continuous but has a sharp corner. G1 means the tangents align but curvature might jump. G2 means both tangent direction and magnitude of curvature are continuous. For freeform surfacing, you need G2 minimum. For a simple 2D path, G1 is often sufficient. The insight nobody tells beginners: you can achieve G2 with fewer control points than you think. A well-placed cubic with three control points per span gives G2 naturally. Adding more points without adjusting the underlying polynomial structure just adds redundancy, not smoothness. My go-to workflow for curve fitting. I start with the minimum number of control points that approximates the shape I need. Two or three spans. Then I check the curvature plot. If it's smooth and monotonic where it should be, I'm done. If there are wiggles, I add control points only at the wiggle locations, not uniformly along the curve. Uniform refinement is lazy and usually makes things worse. I also avoid letting my software's auto-fit tool run unchecked — it tends to overfit noise in the data. I set a tolerance threshold before running it, something like 0.01mm for precision work.

The Downside Nobody Advertises

Splines are not a universal solution. They fail in three scenarios that come up more often than you'd expect. First, high-degree splines. A single polynomial above degree five becomes numerically unstable. The condition number of the basis matrix grows exponentially. Oscillations appear at the ends of the interval — this is Runge's phenomenon, and it's not a bug in your software, it's mathematics. Keep your spline degree at three unless you have a very good reason not to. Three is the sweet spot for continuity, computational cost, and numerical stability combined. Second, periodic shapes with non-periodic boundary conditions. If you're modeling something like a ring or a torus section, you need a closed periodic spline. Most tools handle this with a single extra operation, but if you forget it, the curve will have a visible cusp or discontinuity at the closure point. I once spent a full day debugging a surface that had a tiny but fatal flaw at the seam. The control polygon looked perfect. The issue was purely in how the first and last spans were connected.

Third, data with uneven sampling density. Splines assume a continuous parameter domain. If your input points are clustered in one region and sparse in another, the resulting curve will be over-refined where you don't need it and under-refined where you do. The fix is reparameterization — redistributing the points by chord length or centripetal parameterization before fitting. Without this step, you're fitting a curve to raw data, and raw data rarely respects the assumptions the algorithm makes. If you're working in a domain where these failure modes are likely — high-precision machining, aerodynamic surfaces, or anything involving tight tolerances — consider whether a Bézier approach or even a simple polyline approximation might be more appropriate. Splines are powerful but they carry assumptions that, when violated, produce subtle failures that are expensive to diagnose. The short version: know your continuity requirements before you start, clamp your knots, keep degree at three, and always check the curvature plot. Everything else is detail work.