> allowing them to move from one solution to the other smoothly during the optimization step Yes, and they are making this explicit in the paper. However, whe

> allowing them to move from one solution to the other smoothly during the optimization step Yes, and they are making this explicit in the paper. However, where my intuition "doesn't believe them" is because: - The problem has *physical* local extrema (e.g. two 3D circles of identical radius and axis when dipped in a soap solution can yield either a catenoid or two flat disks. If the circles are close enough to one another, only the former is a true global extremum while the latter is a local one). I can't visualize a transformation of the problem that would somehow simply rub out those local extrema: once your optimizer gets close to the two circles, I'm having a hard time seeing how it can "jump" back to the catenoid, *especially* if the method is local. - the border can have an arbitrarily complex shape. That's a *huge* space to design from for a pathological border that will yield a whole zoo of local extrema. I'm willing to believe they found a way, but I need to learn about differential forms before I can even try to understand the paper. Also they claim to have published the source code to their algorithm, but upon download, it's turns out to be a binary turd that can only be interpreted by a proprietary package (Houdini) that coredumps on you unless it runs an OpenCL capable system. Not exactly what I'd call easy to read code.