Numerical Differentiation of SDEs

• Numerical differentiation of SDEs is the process of estimating gradients of expected outcomes with respect to system parameters, addressing challenges from noise and discretization errors.
• The discretize-optimize approach discretizes SDEs into finite Markov chains then applies sensitivity analysis, while the optimize-discretize method differentiates a continuous-time adjoint equation.
• Proper matching of the numerical scheme (e.g., Euler–Maruyama, Heun) with the SDE interpretation (Itô, Stratonovich) is critical for unbiased gradient estimates in applications like financial modeling.

Numerical differentiation of stochastic differential equations (SDEs) entails the computation of parameter sensitivities—gradients of an expected terminal cost functional with respect to system parameters—when analytic solutions are unavailable. Given an SDE parameterized by $\theta$ with terminal objective $J(\theta) = \mathbb{E}[\Phi(X(T;\theta))]$, the problem reduces to estimating $\nabla_\theta J(\theta)$. Stochasticity and discretization introduce significant challenges compared to deterministic ordinary differential equations (ODEs), requiring specialized numerical approaches that account for the interplay between noise, discretization schemes, and differentiation. The core strategies consist of the discretize-optimize and optimize-discretize paradigms, each with distinct theoretical and practical implications for Itô and Stratonovich SDEs (Leburu et al., 13 Jan 2026).

1. Formulation and SDE Conventions

Consider the SDE

$$dX(t) = f(t, X(t), \theta)\,dt + g(t, X(t), \theta)\,dW(t), \quad X(0) = x_0,$$

where $W(t)$ is an $m$-dimensional Brownian motion, $\theta$ are parameters which may include $x_0$, $f$ is the drift, and $g$ the diffusion. The terminal cost $\Phi:\mathbb{R}^d \to \mathbb{R}$ defines the objective $J(\theta) = \mathbb{E}[\Phi(X(T;\theta))]$. Two stochastic calculus conventions are employed:

• Itô SDE: Differential as above, interpreted in the Itô sense.
• Stratonovich SDE: The $dW(t)$ increment is replaced by $\circ dW(t)$, with calculus rules corresponding to the chain rule.

The Itô and Stratonovich forms are linked by a drift correction: $$\hat{f}_i(t,x) = f_i(t,x) - \frac{1}{2} \sum_{j=1}^m\sum_{k=1}^d \frac{\partial g_{ij}}{\partial x_k}(t,x)\cdot g_{kj}(t,x),$$ so that the Itô SDE with $(f,g)$ is equivalent in law to the Stratonovich SDE with $(\hat{f},g)$. Selection of the calculus convention and corresponding numerical discretization profoundly influences the validity of gradient estimators.

2. Discretize-Optimize Approach

The discretize-optimize paradigm involves discretizing the SDE to form a finite-dimensional Markov chain, then differentiating the resulting discrete objective either via forward/reverse mode sensitivity analysis or automatic differentiation. Standard discretizations include:

• Euler–Maruyama (Itô):

$$X_{n+1} = X_n + f(t_n, X_n, \theta)\,\Delta t + g(t_n, X_n, \theta)\,\Delta W_n,$$

for $n=0, \ldots, N-1$ with $\Delta t = T/N$ and $\Delta W_n \sim \mathcal{N}(0, \Delta t I_m)$.

• Heun (Stratonovich):

$$\begin{aligned} \tilde{X}_{n+1} &= X_n + f(t_n, X_n)\Delta t + g(t_n, X_n)\Delta W_n, \ X_{n+1} &= X_n + \frac{1}{2}\big[f(t_n, X_n) + f(t_{n+1}, \tilde{X}_{n+1})\big]\Delta t \ &\phantom{=} + \frac{1}{2}\big[g(t_n, X_n) + g(t_{n+1}, \tilde{X}_{n+1})\big]\Delta W_n. \end{aligned}$$

This scheme is pathwise symmetric and recovers Stratonovich calculus.

Pathwise gradients are obtained by interchanging $\nabla$ and expectation, yielding $$\nabla_\theta J_{\Delta t} = \mathbb{E}[\nabla_\theta \Phi(X_N)]$$. The chain rule over time steps propagates gradients through the sequence of discrete updates.

Forward-mode and reverse-mode (adjoint) sensitivity recursions yield, respectively: $$\begin{aligned} A_n &:= \frac{\partial X_{n+1}}{\partial X_n} = I_d + \partial_x f\,\Delta t + \partial_x g\,\Delta W_n,\ B_n &:= \frac{\partial X_{n+1}}{\partial \theta} = \partial_\theta f\,\Delta t + \partial_\theta g\,\Delta W_n, \end{aligned}$$ with appropriate Jacobian products over time. The discrete adjoint recursion iterates backward: $$\begin{aligned} p_N &= \partial_x\Phi(X_N)^\top,\ p_n &= A_n^\top p_{n+1},\ q_n &= q_{n+1} + B_n^\top p_{n+1}. \end{aligned}$$ The terminal pathwise gradients for $x_0$ and $\theta$ are $p_0$ and $q_0$ respectively. For Monte Carlo approximation, per-sample gradients are averaged over $M$ simulated trajectories.

3. Optimize-Discretize Approach

The optimize-discretize method first derives a continuous-time backward equation reflecting the parametric sensitivities, then discretizes this adjoint SDE. For the Stratonovich SDE, the continuous-time adjoint $p(t)$ satisfies: $$dp(t) = -[\partial_x f(t,X(t))]^\top p(t)\,dt - \sum_{j=1}^m [\partial_x g_{:j}(t,X(t))]^\top p(t) \circ dW^j(t), \qquad p(T) = \partial_x\Phi(X(T))^\top.$$ The gradient $\nabla_{x_0} J$ is given by $\mathbb{E}[p(0)]$.

For Itô SDEs, conversion to Stratonovich via drift correction is required before differentiation, or alternatively, the continuous adjoint equation may be written in equivalent Itô form. Naive backward-Euler discretization of the continuous adjoint may fail to converge to the correct gradient unless the diffusion Jacobian $\partial_x g$ is state-independent. Counterexamples include SDEs with non-constant diffusion coefficients, where a bias results unless further correction is applied.

Discrete adjoint recursions for the Heun scheme converge to the continuous Stratonovich adjoint as $\Delta t \to 0$, while those for Euler–Maruyama converge only in special cases.

4. Agreement and Divergence of Methods

The two approaches coincide or diverge based on the underlying SDE structure, as follows:

Setting	Agreement Condition	Outcome
Deterministic ODE	$g \equiv 0$	Methods coincide
Stratonovich SDE	Pathwise symmetric (Heun) discretization	Methods coincide
Itô SDE	$\partial_x g$ constant (e.g., Black–Scholes)	Methods coincide
Itô SDE	$\partial_x g$ state-dependent (e.g., CEV model)	Methods diverge, bias occurs

In ODEs and Stratonovich SDEs (or Itô with constant Jacobian noise), either method recovers correct gradients in the $\Delta t \to 0$ limit. For generic Itô SDEs with state-dependent diffusion, optimize-discretize is biased unless drift correction and suitable discretization are applied. This subtlety is critical in financial and physical modeling where models often feature non-constant diffusion terms.

5. Algorithmic Implementation

A summary of algorithmic workflows for each approach is as follows:

Approach	Steps	Notes
Discretize-Optimize (EM)	Simulate forward Euler–Maruyama;	Gradient is exact for discrete obj;
	Run backward adjoint recursion	$\mathcal{O}(\sqrt{\Delta t})$ error
Discretize-Optimize (Heun)	Simulate forward Heun; backward adjoint with higher-order terms	Converges to Stratonovich adjoint
Optimize-Discretize (Stratonovich)	Simulate forward path; backward SDE integration	No trajectory storage required if path reversal is used

Per-path unbiasedness is maintained by reusing the same Brownian increments in forward and backward simulations. Memory complexity to store forward trajectories is $\mathcal{O}(N\cdot d)$. For high-dimensional or long-horizon problems, checkpointing provides a trade-off between memory and computation.

6. Representative Models and Case Studies

Two illustrative examples highlight the nuances:

• Black–Scholes model (Itô): $dS = rS\,dt + \sigma S\,dW$. Both discretize-optimize and optimize-discretize (naive Itô adjoint) yield unbiased gradient (Delta) estimates, as $\partial_S g$ is constant. Monte Carlo average converges at rate $\mathcal{O}(\sqrt{\Delta t})$ (Leburu et al., 13 Jan 2026).
• CEV model (Itô, state-dependent diffusion): $dS = rS\,dt + \sigma S^\beta\,dW$, $\beta \neq 1$. Discretize-optimize Euler–Maruyama adjoint yields unbiased $\Delta \approx 0.64$, but naive optimize-discretize backward-Euler produces a large bias ($\approx 1.36$) with heavy-tailed errors. Conversion to Stratonovich and Heun discretization rectifies the bias, with consistent unbiased estimates matching discretize-optimize.

This demonstrates the necessity of matching the calculus convention, discretization scheme, and differentiation method for reliable sensitivity estimates, especially in models with non-trivial diffusion structure.

7. Practical Issues and Recommendations

Key practical considerations for numerical differentiation of SDEs include:

• Always employ identical Brownian trajectories for forward and backward passes to ensure unbiased Monte Carlo gradient estimates per sample.
• For problems with large $N$ or $d$, checkpointing can alleviate memory constraints at the cost of recomputation.
• Stratonovich discretizations (e.g., Heun) possess pathwise reversibility, facilitating backward reconstruction without storing full trajectories.
• For pure parameter sensitivity (such as financial Greeks), parameters can be embedded as constant states with adjoint recursion computed simultaneously.
• Pathwise differentiation can be extended to cost functionals including running costs by augmenting the state with an integral and adjusting the terminal condition.
• Discretize-optimize is robust across settings: it directly yields the exact gradient for the chosen discrete scheme, with convergence guarantees as the discretization refines, provided the scheme matches the calculus (Itô or Stratonovich). Optimize-discretize is delicate—unless appropriately corrected, it may yield fundamentally incorrect gradients and should be applied with understanding of its structural limitations.

These considerations dictate the appropriate methodology based on accuracy, memory, analysis requirements, and the analytic structure of the SDE (Leburu et al., 13 Jan 2026).