Backward stochastic differential equations (BSDEs) are mathematical tools used in finance, control theory, and probability to model systems where you know the final outcome but need to work backward in time to understand the full dynamics. Solving them numerically often means replacing the continuous randomness of Brownian motion with a simpler, finite set of possible moves at each time step. This paper studies how well such discrete approximations can capture two desirable properties simultaneously: accurately tracking the solution's value and correctly representing the randomness driving the system through a clean decomposition of uncertainty.
The central finding is a fundamental tension between two goals. If you design your finite set of random moves to have a vanishing "third moment" (roughly, the moves are symmetric in a strong sense), you get first-order accuracy in approximating the solution. But in two or more dimensions, that same symmetry condition turns out to be incompatible with perfectly representing the underlying randomness without any leftover noise. There is always a residual martingale component that cannot be captured, and its size is controlled by a precise geometric quantity tied to how well your discrete distribution approximates a continuous sphere. The best possible distributions for minimizing this error are those related to objects called spherical designs, which are configurations of points that integrate polynomials exactly up to a certain degree.
The paper makes this tradeoff precise and quantitative. The same mathematical object, a third-order tensor describing the asymmetry of the one-step distribution, simultaneously governs how accurate the approximation is and how large the unavoidable representation error is. For a specific simple example, the regular simplex rule, the authors compute everything explicitly: the exact rate at which the residual noise accumulates, and how a mismatch in the cubic structure of the terminal condition translates into a concrete weak approximation error. The upshot is a kind of uncertainty principle for BSDE numerics: you cannot simultaneously achieve perfect symmetry for accuracy and perfect representation of randomness with any finite collection of moves in higher dimensions.