The paper tackles a mathematical problem involving "backward stochastic variational inequalities" (BSVIs) in two dimensions. These are equations that describe systems evolving randomly over time, but solved backward from a final condition rather than forward from an initial one. The key feature here is that the solution is constrained to stay inside a curved, convex region, and whenever it hits the boundary, it gets reflected back inward. The twist is that the reflection does not happen straight inward (symmetrically) but at an angle that rotates and varies with time. The authors work in a very general probabilistic setting that allows for sudden jumps in the randomness, not just smooth continuous fluctuations.
The main technical challenge is that the non-symmetric (rotated) reflection direction introduces an extra term along the boundary that breaks the standard mathematical tools used to prove such equations have unique solutions. Standard approaches rely on a kind of stability property called monotonicity, which the tilted reflection destroys. The authors invent a clever workaround: they construct a specially designed mathematical object called a symmetric two-point kernel with position-dependent coefficients. This kernel is engineered so that its behavior at the boundary precisely cancels out the troublesome extra term. They then use a refined probabilistic estimate involving exponential weighting to handle the remaining errors coming from both the curvature of the problem and the jumps in the random process.
Building on this existence and uniqueness result, the authors also study a related control problem: what if someone can choose how much the reflection angle rotates over time, subject to limits on how fast it can change? They show that an optimal choice of this rotation strategy always exists within a natural class of allowable controls. The work has potential relevance wherever constrained stochastic systems arise, such as in mathematical finance (for pricing under constraints), queuing networks, or other applied areas where random dynamics must respect physical or financial boundaries.