In computational geometry, robotics, and machine learning, researchers often rely on a powerful simplifying assumption: that the world is made of convex shapes. A set is convex if, for any two points within it, the straight line connecting them also lies entirely within the set. This property is the "easy mode" of mathematics. It ensures that local optimizations lead to global solutions, that paths do not get trapped in dead ends, and that complex boundaries can be described by simple linear inequalities.
However, as recent research across multiple domains shows, while the convexity assumption makes problems tractable, it often acts as a bottleneck that prevents algorithms from generalizing to the messy, non-convex reality of the physical world.
The Appeal of Convexity
The primary reason researchers like Prasad Chalasani or Robbins et al. start with convex polygons is that they allow for predictable geometric behavior. In a convex environment, the "visibility" between points is well-behaved. For example, Robbins et al. developed an efficient $O(n^2 \log n)$ algorithm for curvature-constrained shortest paths, but they explicitly noted that extending this to non-convex polygons remains a significant challenge because the path classification becomes far more complex when walls can "hide" parts of the space.
This simplification extends to how we represent obstacles. Chalasani and Beul et al. highlight that modeling obstacles as axis-parallel rectangles or hyperrectangles—a specific type of convex set—dramatically reduces the math needed for collision avoidance. By assuming convexity, planners can use tools like "coresets" to bound obstacles with simple ellipsoids, as demonstrated by Tukan et al., which allows them to prune the search space efficiently.
Enforcing Safety via Projections
In the realm of deep learning and control, convexity is not just a simplification; it is a guarantee. When a neural network must control a self-driving car or a robot arm, its outputs must stay within "safe" limits.
Architectures like ConstraintNet, proposed by Brosowsky et al., and $\Pi$net, developed by Grontas et al., enforce these hard constraints by projecting the network's raw output onto a feasible set. This works because, for a closed convex set, there is always a unique "closest point" to any given output. Grontas et al. use orthogonal projection layers to ensure safety, but they acknowledge that this mechanism is currently limited to convex sets. If the safe region were non-convex (like a U-shaped room), a simple projection might jump discontinuously between different parts of the set, leading to unstable behavior.
Similarly, in dynamical systems, the authors of Projected incrementally scattering passive systems on closed convex sets show that passivity and stability can be maintained when a system is restricted to a convex set. Moving to non-convex constraints would require "fundamentally different mathematical approaches" to ensure the system doesn't become unpredictable at the boundaries.
The Computational Cost of Realism
When we move away from the convexity assumption, computational complexity often explodes. This is clearly visible in pathfinding and volume estimation:
- Pathfinding: Mitchell et al. show that while finding paths in rectilinear (axis-parallel) domains is efficient, generalizing to arbitrary angles or non-convex obstacles can turn a simple problem into a "3SUM-hard" one, where sub-quadratic solutions are difficult to find.
- Volume Estimation: Calculating the volume of a high-dimensional shape is essential for Bayesian statistics and robotics. Ge et al. and Chevallier et al. provide fast algorithms for this, but their methods are strictly bound to convex polytopes. Non-convex shapes require decomposing the space into many convex sub-regions, a process that is often as difficult as the original problem.
- Numerical Analysis: In finite element methods, which are used to simulate physics, the construction of basis functions often assumes convex polygons or polyhedra. Gillette et al. note that extending these frameworks to non-convex shapes is an open challenge because the standard "barycentric coordinates" used to interpolate values across a shape don't behave well when the shape has "dents."
Breaking the Assumption
The frontier of this research area lies in finding ways to relax the convexity constraint without losing theoretical guarantees. Zhiqiang Cai and Luke Olson have explored how numerical error estimators and interface solvers—traditionally reliant on convex domains—can be adapted for more complex geometries. In tomography, Kolehmainen and Ola proved that while strictly convex domains allow for unique recovery of internal properties, it remains an open question if these principles hold for the concave shapes common in biological applications.
Ultimately, the convexity assumption is a bridge. It has allowed researchers to build the first generation of optimal navigation algorithms, safe neural networks, and efficient high-dimensional counters. The next challenge is building the mathematical tools to step off that bridge into the non-convex world.
Key Foundational Work
An Online Algorithm for Improving Performance in Navigation (2000) — contributed 1 concept elements to this direction.
An on-line algorithm for improving performance in navigation (1994) — contributed 1 concept elements to this direction.