When a stochastic partial differential equation (SPDE, a PDE driven by random forcing, such as a heat equation with noise that multiplies the solution) is simulated with a time-stepping scheme, there are two ways to judge the error. Strong error compares individual simulated paths with the true paths. Weak error compares only averages of smooth quantities at the final time. A decade-long line of work culminating in Weak Convergence Rates for Euler-Type Approximations of Semilinear Stochastic Evolution Equations with Nonlinear Diffusion Coefficients showed that the weak error of Euler-type schemes decays twice as fast as the strong error. For the parabolic Anderson model driven by space-time white noise, the weak rate is nearly $N^{-1/2}$ in the number of time steps $N$, while the strong rate is $N^{-1/4}$. That guarantee covers only expectations at a fixed terminal time. Many uses of a simulator involve a decision-maker who watches the solution evolve and chooses when to act: stopping a process, exercising an option on a spatially distributed quantity, triggering an intervention. For those uses the relevant quantity is an optimal stopping value, and the adapted weak topology literature (A note on the adapted weak topology in discrete time, The Wasserstein space of stochastic processes) shows that terminal-time closeness does not imply closeness of such values.
This note proposes to determine whether the doubling of the rate survives when the test is an optimal stopping problem rather than a terminal expectation. The natural tool on the transport side is the adapted Wasserstein distance, which compares two stochastic processes using only couplings that never look into the future. It controls stopping values directly, but the obvious coupling of scheme and true solution (driving both with the same noise) only yields the strong rate $N^{-1/4}$. Our claim is that for smooth rewards, stopping values of Euler-type SPDE schemes converge at the weak rate $N^{-(1-\gamma-\epsilon)}$, where $\gamma$ is the noise-regularity exponent. The route to this is the mild Itô formula with stopping times, the operator-theoretic chain rule for SPDEs that can be evaluated at random times. We also claim that the adapted Wasserstein distance itself is stuck at the strong rate. If both hold, a "smooth" adapted error is the correct way to certify SPDE simulators for sequential decisions, and the Lipschitz-cost adapted distance is too coarse by a factor of two in the exponent.
Core Claim
Let $X$ be the mild solution of a semilinear SPDE $dX_t = (AX_t + F(X_t))\,dt + B(X_t)\,dW_t$ on a Hilbert or type-2 UMD Banach space $E$, with $A$ the unbounded operator (for example the Laplacian) and $W$ the driving noise. Let $Y^N$ be its linear-implicit or exponential Euler approximation with $N$ steps. For a reward $g\colon[0,T]\times E\to\mathbb{R}$ and a set $\mathcal{T}$ of admissible exercise times, define the stopping value
$$V_g(X) = \sup_{\tau \in \mathcal{T}} \mathbb{E}\big[g(\tau, X_\tau)\big],$$
where the supremum is over stopping times of the process's own filtration. Define the adapted weak error of order $k$ as $\sup\{|V_g(X)-V_g(Y^N)| : \|g\|_{C_b^k}\le 1\}$. We conjecture, and aim to prove, that
$$\sup_{\|g\|_{C_b^k}\le 1} |V_g(X) - V_g(Y^N)| \le C\, N^{-(1-\gamma-\epsilon)},$$
first for Bermudan exercise at a fixed finite set of $M$ dates, and then uniformly as $M$ grows to continuous (American) exercise. The second, negative claim is that $\mathcal{AW}_1(X, Y^N)$, the adapted Wasserstein distance with a Lipschitz cost on paths, decays no faster than the strong rate $N^{-1/4}$ for the parabolic Anderson model. If so, a proof through adapted Wasserstein stability cannot reach the weak rate.
The conjecture should not be read as obviously true. The Snell envelope is the value of stopping optimally from each state onward, built backward in time by taking the maximum of "stop now" and "continue". That maximum creates kinks, so the backward value function is not five times differentiable, which is what the Jentzen–Kurniawan argument requires. This is exactly the non-smooth test function obstacle that the file lists as the main open problem of the mild Itô program. The honest range of outcomes for hard maxima is therefore a rate somewhere between $1/4$ and $1/2$. Pinning down which, and why, is the contribution.
Research Program
The analysis reuses the machinery of Weak Convergence Rates for Euler-Type Approximations of Semilinear Stochastic Evolution Equations with Nonlinear Diffusion Coefficients and makes one change: the terminal Kolmogorov function is replaced by a backward dynamic-programming recursion. That paper supplies three ingredients. The first is the mild Itô formula with stopping times from On the mild Itô formula in Banach spaces. The second is the continuous interpolation ("semilinear integrated counterpart") of the Euler scheme. The third is Fréchet-derivative bounds for the backward Kolmogorov equation $u(t,x)=\mathbb{E}[\phi(X^x_{T-t})]$, where $u$ is the expected final payoff started from state $x$ at time $t$.
The work proceeds in three stages. In the first, Bermudan exercise with a fixed number $M$ of dates, we replace the hard maximum by a soft maximum at temperature $\eta$. This gives an entropically regularized stopping value whose backward recursion keeps $C_b^5$ regularity, with constants that grow like a negative power of $\eta$. Between exercise dates the recursion is a Kolmogorov problem, so the existing weak error estimate applies on each interval. The mild Itô formula with stopping times is what lets the error expansion be evaluated at the random exercise time instead of at $T$, and it provides localization for coefficients that are only locally Lipschitz. Balancing the smoothing bias, which is of order $\eta$, against the derivative growth gives a rate for hard stopping. The key technical question is whether the analytic smoothing of the semigroup $S(t)=e^{tA}$ over each inter-date interval restores enough regularity to the continuation value that the soft maximum can be removed at no cost to the rate. This is plausible because the continuation value is a conditional expectation over a strictly positive time interval, where the semigroup and noise act together to smooth.
In the second stage, the American limit, we let $M$ grow with $N$ and track how constants depend on $M$. Here the stopping-time version of the mild Itô formula is used in its full form, applied to the regularized value function along the path up to a genuinely continuous stopping time. This is also the stage where the curse of the time horizon noted for adapted transport bounds will show up if it is real.
In the third stage we prove or refute the negative claim. The synchronous coupling, which drives the scheme and the true solution with the same noise, is bicausal (it never uses future information in either direction), so the upper bound $\mathcal{AW}_1(X,Y^N)\le C N^{-1/4}$ is immediate. For the lower bound, we will try to construct an adapted Lipschitz functional of the path whose expectation differs by order $N^{-1/4}$. The difference between increments of the scheme and increments of the true solution at the finest scale is a natural candidate. Numerically, we will estimate $\mathcal{AW}_1$ directly. A proven gap between the adapted weak error and $\mathcal{AW}_1$ is the precise sense in which the two research directions meet: the first can be proven to reach the weak rate through mild Itô calculus, while the second can only be controlled through strong error.
Baselines
The naive reference point is the strong-rate bound obtained through adapted transport. Adapted Wasserstein closeness controls stopping values with a constant set by the Lipschitz constant of $g$, as formalized in Bounding adapted Wasserstein metrics. The synchronous coupling bounds $\mathcal{AW}_1$ by the strong error. For the parabolic Anderson model the strong rate is $1/4$, as reported in Weak Convergence Rates for Euler-Type Approximations of Semilinear Stochastic Evolution Equations with Nonlinear Diffusion Coefficients. For the truncated schemes on stochastic Ginzburg–Landau, Strong convergence rates for nonlinearity-truncated Euler-type approximations of stochastic Ginzburg–Landau equations reports an empirical strong root-mean-square error slope of approximately $1/4$. Beating this baseline means an empirical slope of the stopping-value error clearly steeper than $1/4$, together with a proof of a rate above $1/4$. Both will be measured on the same reference solutions and exercise dates.
The strongest existing method is the terminal weak error analysis of Weak Convergence Rates for Euler-Type Approximations of Semilinear Stochastic Evolution Equations with Nonlinear Diffusion Coefficients. It proves a weak rate of "nearly $1/2$, which is exactly double the known strong convergence rate of $1/4$" for the parabolic Anderson model with $C_b^5$ test functions. Weak convergence rates for temporal numerical approximations of stochastic wave equations with multiplicative noise proves the analogous doubling for the hyperbolic model, without a reported constant. We will run the terminal-time version of every experiment (the reward evaluated only at $T$, with no stopping) alongside the stopping version, using identical grids, reference solutions and Monte Carlo budgets. The terminal curve is therefore a ceiling rather than a competitor. Matching its slope with stopping enabled is the success criterion.
Three ablations isolate the key ingredients. Varying the soft-maximum temperature $\eta$, including hard stopping at $\eta = 0$, tests whether semigroup smoothing makes the regularization unnecessary. Replacing $\mathcal{AW}_1$ by the classical path Wasserstein distance $\mathcal{W}_1$ on the same discretized paths tests whether the adapted constraint matters at all at these resolutions. Growing the number of exercise dates $M$ with $N$ tests uniformity in the American limit.
Why not just…?
Why not just bound stopping values by the adapted Wasserstein distance and stop there? Because that route is capped by the best coupling one can construct, and the only coupling available in general is the synchronous one, which gives rate $1/4$. The alternative smoothing bridge $\mathcal{AW}_1\le C\sqrt{\mathcal{W}_1}$ from Bounding adapted Wasserstein metrics would, even if extended to infinite dimensions, lose a further square root. It also assumes Lipschitz transition kernels in finite-dimensional discrete time, which space-time white-noise SPDEs do not obviously satisfy. Stage three directly tests whether this route is fundamentally slower. If $\mathcal{AW}_1$ turned out to decay at $N^{-1/2}$ empirically, this alternative would be a genuine threat, and we would redirect the program toward proving a sharper adapted coupling.
Why not just apply the terminal weak error bound at each exercise date? Closeness of the marginal laws at each date does not control a supremum over stopping rules, because the optimal rule depends on the path. The Game A versus Game B example and the results of Causal transport in discrete time and applications show that matching laws at each time can coexist with different stopping values. What does work is backward induction with a weak error estimate on each inter-date interval. That is essentially the first stage of our program, and we do not claim the Bermudan fixed-$M$ smooth case is deep. The substance is removing the soft maximum and obtaining uniformity in $M$, which per-date terminal bounds cannot provide.
Why not just estimate stopping values by simulation and regression and report the empirical slope? Without theory, Monte Carlo and regression bias make it very hard to tell a slope of $1/2$ from one of $0.4$ at feasible $N$. An empirical slope also cannot tell us whether a result transfers to other equations. We will run exactly this experiment as a check, but paired with a theorem and with the soft-maximum ablation, which predicts how the slope should change with $\eta$.
Proposed Experiments
The benchmarks are the three SPDEs in the file, all in one space dimension. The first is the parabolic Anderson model $dX_t=\Delta X_t\,dt + X_t\,dW_t$ with space-time white noise ($\gamma=1/2$). The second is the stochastic Ginzburg–Landau equation with cubic drift, simulated with the truncated schemes of Strong convergence rates for nonlinearity-truncated Euler-type approximations of stochastic Ginzburg–Landau equations. The third is the hyperbolic Anderson model. The reference solutions are high-resolution "overkill" simulations, following the MATLAB protocol of that paper.
Rewards are smooth functionals such as $g(t,x)=e^{-rt}\psi(\langle x, e_1\rangle,\|P_K x\|^2)$, where $\psi$ is a smooth function, $e_1$ is the first eigenfunction of the Laplacian, $P_K$ projects onto the first $K$ Fourier modes, and $r$ is a discount rate. The exercise dates are $M\in\{4,16,64\}$. Stopping values are computed by backward regression on a fixed feature basis. The same basis and the same random numbers are used for the scheme and the reference, so that regression bias largely cancels in the difference. The regression error itself is monitored by varying basis size.
Adapted Wasserstein distances between scheme and reference are computed with the adapted Sinkhorn algorithm of Computational methods for adapted optimal transport, applied to paths projected onto a few Fourier modes and a few time slices. We keep branching small, since that paper reports exact backward induction to be competitive only for "very small numbers of branches ($n_b = 10, 25$)". We report $\mathcal{W}_1$ on the same projected paths. No reported numerical result was found on Lacuna for adapted distances on SPDE paths, so these curves are new measurements rather than comparisons with published values.
The decisive check is a log-log plot of stopping-value error against $N$ for the parabolic Anderson model with hard stopping. A slope near $1/2$, matching the terminal curve, supports the main claim. A slope near $1/4$ means that stopping destroys the doubling. An intermediate slope that moves with $\eta$ indicates that regularity is the bottleneck. The second check is whether the $\mathcal{AW}_1$ curve stays near slope $1/4$ while the adapted weak error reaches $1/2$. That gap is the note's central claim that smooth adapted error, not the Lipschitz adapted distance, is the right yardstick.
As a secondary application, the same stopping-value test will be applied to the Neural SPDE surrogate of Neural Stochastic PDEs: Resolution-Invariant Learning of Continuous Spatiotemporal Dynamics. On Ginzburg–Landau it reports a relative $L^2$ error of 0.006, against 0.027 for a Fourier neural operator. The question is whether a small $L^2$ error translates into accurate stopping values, or whether the surrogate reproduces paths while distorting what can be learned from observing them over time.
References
- Weak Convergence Rates for Euler-Type Approximations of Semilinear Stochastic Evolution Equations with Nonlinear Diffusion Coefficients
- A note on the adapted weak topology in discrete time
- The Wasserstein space of stochastic processes
- On the mild Itô formula in Banach spaces
- Bounding adapted Wasserstein metrics
- Strong convergence rates for nonlinearity-truncated Euler-type approximations of stochastic Ginzburg–Landau equations
- Weak convergence rates for temporal numerical approximations of stochastic wave equations with multiplicative noise
- Causal Transport in Discrete Time and Applications
- Computational methods for adapted optimal transport
- Neural Stochastic PDEs: Resolution-Invariant Learning of Continuous Spatiotemporal Dynamics