Backward stochastic differential equations (BSDEs) are equations that run backwards in time and are used widely in finance, control theory, and probability. This paper studies a version where the equation is driven not by ordinary noise but by a "rough path," a mathematical object that captures highly irregular signals more general than standard Brownian motion. The central question is: if you slightly change the rough path driving the equation, how does the solution change? Answering this carefully requires sensitivity analysis, which for ordinary equations would just mean taking a derivative. Here the difficulty is that rough paths do not live in a simple flat vector space, so the usual notion of a derivative does not directly apply.
The paper introduces a rigorous framework for defining what a derivative means in this setting. The key insight is to work with a "tangent structure" that keeps track of both a first-level direction (roughly, the direction you perturb the path) and a compatible second-level correction (needed because rough paths carry extra geometric information beyond their basic trajectory). The authors prove that, under reasonable conditions on the equation, the solution responds to small perturbations in a way that is well approximated by a continuous linear map acting on these two-level directions. This is called intrinsic tangential Hadamard differentiability, meaning the derivative exists and is stable, but only along carefully chosen sequences of perturbations rather than in all directions at once.
The technical work involves several layers of analysis: expanding the flow of the equation to high order, applying a classical transformation to simplify the generator of the equation, and then using delicate estimates from the theory of bounded mean oscillation to control errors uniformly. The upshot is a precise and coordinate-independent statement: the sensitivity of the solution is captured by a well-defined linear object that does not depend on arbitrary choices made in representing the rough path. This advances the mathematical foundations needed to do rigorous calculus on spaces of rough paths, with potential applications wherever rough-path-driven models need to be calibrated or optimized.