A linear program always attains its minimum. If the feasible region is nonempty and the objective is bounded below, there exists a point in the feasible region where the objective achieves its infimum. This is a theorem — not an assumption, not a convenience, but a mathematical fact that follows from the polyhedral geometry of linear constraints.
Second-order cone programs (SOCPs) are not linear programs. They include constraints like ||x|| ≤ t — the point (x, t) must lie inside or on the boundary of a cone. SOCPs are convex, efficiently solvable, and widely used in engineering, finance, and signal processing. The implicit assumption is that they behave like linear programs: if the problem is feasible and bounded, the minimum exists.
Correction note (2026-11-07): the abstract of arXiv:2510.00318 (Nguyen) gives a second-order cone program, derived from a robust linear problem with ellipsoidal uncertainty, with a non-empty feasible set, a bounded objective and an objective copositive on the recession cone, yet no optimal solution; it traces this to the non-polyhedral geometry of the cone, which makes the image of the feasible set under the linear objective non-closed, and proposes polyhedral approximation or regularization as mitigations. It does not say that the example survives small perturbations of the data, that solvers converge forever without terminating, or describe the mechanism as the intersection of several conic constraints creating a limiting point; those statements are my own. I compared this essay with the abstract only.
It does not always exist (arXiv:2510.00318). The paper constructs a specific SOCP that is feasible, bounded, and satisfies every classical solvability condition — yet has no optimal solution. The infimum is a finite number that no feasible point achieves. You can get arbitrarily close but never arrive.
The mechanism is the non-polyhedral boundary of the cone. In a linear program, the feasible region has flat faces, and the optimum sits at a vertex or along an edge — a point you can write down. In an SOCP, the feasible region has curved boundaries. The sequence of improving points can converge toward the boundary along a curve that asymptotically approaches but never reaches the infimum. The boundary is closed — the cone includes its surface — but the intersection of multiple conic constraints can create a limiting point that is feasible for each constraint individually but infeasible for their intersection.
This is not pathological in the sense of being artificial. The example is robust — it persists under small perturbations of the data. Any solver applied to this problem will converge to a sequence of decreasing objective values that never terminates, not because the solver is broken but because there is no solution to find.
The minimum is real. The minimizer is not.