Error Propagation Is Traced Backward in Discrete-Time Stochastic Control

If a decision-making algorithm makes a small mistake today, how far can that error travel by the end of the problem? This paper studies that question in discrete-time stochastic optimal control, where choices are made step by step under uncertainty. The authors estimate the value function — the number that tells you how good each choice is — by combining kernel ridge regression, a nonparametric regression method in a reproducing kernel Hilbert space (RKHS), with Monte Carlo subsampling for the continuation value, meaning the payoff expected if you keep going. They then build an error decomposition and control the error terms at each time step. The central result is an analysis of how those errors propagate backward in time, from maturity to the initial stage, which the paper says is relatively underexplored in stochastic control. The same framework is then applied to American option pricing, showing how a theory of backward error can speak directly to a major financial problem.

A mistake at the last step can still change the first move. That is the unsettling idea at the heart of this paper. In discrete-time stochastic optimal control, a controller makes choices in stages under uncertainty. Each stage uses a value function, a number that says how good a choice is if you keep going. The surprise is not that errors exist. It is that they can travel backward through time. A flaw near maturity can shape the estimate at the start. That matters for any planner that must act step by step, from finance to robotics.

Why a late error can haunt the first decision

The paper studies this backward drift in a general dynamic programming frame. Dynamic programming breaks a big decision into smaller ones. Each stage depends on the next stage. The authors build a natural error split for the estimated value function. Then they control the error pieces one time step at a time. Their main focus is not a new task. It is the path that error takes. The analysis shows how uncertainty at one stage can feed the next stage back, all the way from maturity to the initial stage. That backward path has been less studied in stochastic control, even though it sits at the core of the method.

The estimate is built in two layers

The value function comes from two tools working together. The first is kernel ridge regression, or KRR, a way to fit a smooth pattern from data without forcing a fixed shape. It runs inside a reproducing kernel Hilbert space, or RKHS, which is a math space that lets the fit stay flexible. The second tool is Monte Carlo subsampling, which means drawing random samples to estimate the continuation value, the payoff you expect if you keep going. One layer learns the shape. The other layer estimates what comes next. Together they give a stepwise estimate that the paper can track through time.

Why this matters beyond the math

American options give the paper a concrete home. An American option is a contract that can be exercised before the final date. That makes timing part of the value. The same backward logic that tracks control error also fits this pricing problem. In practice, that means the analysis does not stay in the abstract. It speaks to a setting where the best choice today depends on what may happen later, and where a small misread at one step can distort the price seen at the start. The method offers a way to ask not only whether the answer is accurate, but where the inaccuracy first enters.

What this points to next

The clearest next test is not a slogan. It is a hard check on the same backward-error idea in other staged problems. The paper already points to discrete-time stochastic control and American option pricing. That suggests a practical question: how stable is the error trail when the same setup meets a different payoff rule or a different time grid? The value of the work lies in that lens. It turns a vague worry about estimation into a step-by-step map. Once you can see where the error begins, you can judge which stage deserves the most care. That is a useful shift for any system that decides by looking one step ahead.