Föllmer’s Approach in Stochastic Finance

• Föllmer’s Approach is a framework that transforms Gaussian samples into target distributions via measure-valued flows, offering efficient and controlled sampling techniques.
• It underpins the Föllmer–Schweizer decomposition, facilitating locally risk-minimizing hedging strategies in incomplete financial markets through explicit integrand solutions.
• The approach supports practical discretization of stochastic integrals with provable error bounds, ensuring reliable implementations in both continuous and discrete-time financial models.

Föllmer’s Approach

Föllmer’s approach encompasses a range of mathematical techniques in probability, stochastic analysis, and financial mathematics. It is most commonly recognized in three primary domains: measure-valued flows for sampling via the Föllmer flow, quadratic hedging and local risk-minimization via the Föllmer–Schweizer decomposition, and related discretization and practical implementations in both continuous and discrete time. Theoretical underpinnings involve stochastic differential equations, semimartingale theory, and optimal transport, with applications extending to Monte Carlo simulation, machine learning (sampling), and mathematical finance (hedging in incomplete markets).

1. Föllmer Flow and Measure-Valued Sampling

The Föllmer flow is a unit-time ODE flow characterized by the evolution of a random variable whose law interpolates between a Gaussian reference and a given target measure. For a probability measure $\nu(dx) = p(x)\,dx$ on $\mathbb{R}^d$ and a Gaussian reference $\gamma^{\mu,\Sigma}$, the preconditioned Föllmer flow is the solution $(X_t)_{t\in[0,1]}$ to the ODE

$$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$

where

$$V(t,x) = \frac{x-\mu + \Sigma S(t,x)}{t},\quad S(t,x) = \nabla_x \ln \int_{\mathbb{R}^d}\varphi^{t y+(1-t)\mu,\, (1-t^2)\Sigma}(x) p(y)\, dy.$$

This flow deterministically transforms samples from the reference into samples from the target measure at time 1. Preconditioning via covariance adaptation (choosing $\Sigma$ matching $\mathrm{Cov}(\nu)$) controls the Lipschitz properties of $V$, improving numerical efficiency. The velocity field can be approximated analytically or via Gaussian Monte Carlo when no closed form is available (Ding et al., 2023).

This approach admits discretization via Euler’s scheme, with numerical error characterized in Wasserstein-2 distance. Non-asymptotic error bounds relate the error to step size, dimension, and the number of Gaussian samples used for velocity approximation. For $K$ time steps and $\mathbb{R}^d$0 Monte Carlo samples, the bound is

$\mathbb{R}^d$1

given appropriate tuning of $\mathbb{R}^d$2 and $\mathbb{R}^d$3 (Ding et al., 2023).

2. The Föllmer–Schweizer Decomposition in Continuous Time

In financial mathematics, the Föllmer–Schweizer (FS) decomposition is central to locally risk-minimizing hedging in incomplete markets. For an asset modeled via an exponential Lévy process $\mathbb{R}^d$4, the discounted price is $\mathbb{R}^d$5. Given a square-integrable contingent claim $\mathbb{R}^d$6, the FS decomposition writes

$\mathbb{R}^d$7

where $\mathbb{R}^d$8 is $\mathbb{R}^d$9-measurable, $\gamma^{\mu,\Sigma}$0 is a predictable integrand (the local risk-minimizing strategy), and $\gamma^{\mu,\Sigma}$1 is a square-integrable martingale orthogonal to the martingale part of $\gamma^{\mu,\Sigma}$2.

For European-type claims $\gamma^{\mu,\Sigma}$3 with suitable integrability and smoothness, the integrand $\gamma^{\mu,\Sigma}$4 has the explicit representation under the minimal martingale measure $\gamma^{\mu,\Sigma}$5:

$\gamma^{\mu,\Sigma}$6

where $\gamma^{\mu,\Sigma}$7 and $\gamma^{\mu,\Sigma}$8 (Thuan, 2020). This explicit solution enables the efficient computation of local risk-minimizing hedges in exponential Lévy settings.

3. Discretization, Jump Correction, and Convergence Analysis

Discretization of the stochastic integral $\gamma^{\mu,\Sigma}$9 is essential for practical implementation, especially in the presence of jumps. The jump-corrected discretization scheme utilizes a deterministic grid augmented with random stopping times capturing large jumps:

• Construct deterministic time-net $(X_t)_{t\in[0,1]}$0.
• Define stopping times $(X_t)_{t\in[0,1]}$1 for large jumps: $(X_t)_{t\in[0,1]}$2, $(X_t)_{t\in[0,1]}$3.
• Use the “frozen” strategy on $(X_t)_{t\in[0,1]}$4, and correct hedges at $(X_t)_{t\in[0,1]}$5 by instantaneously rebalancing.

The error process

$(X_t)_{t\in[0,1]}$6

can be bounded in weighted $(X_t)_{t\in[0,1]}$7 norm with rates depending on the Lévy measure’s small-jump index $(X_t)_{t\in[0,1]}$8, the smoothness of $(X_t)_{t\in[0,1]}$9, and the chosen mesh. For $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$0, the convergence rate is $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$1, while for jump-dominated models it is reduced according to the tail and regularity parameters (Thuan, 2020).

4. Locally Risk-Minimizing Hedging and Structural Default Models

The FS decomposition underpins locally risk-minimizing hedging in models with jump risk and default. In a finite-variation Lévy model for the underlying $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$2, with default time $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$3, a claim $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$4 can be decomposed as

$$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$5

where $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$6 solves an integro-differential equation (PIDE) related to the generator $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$7, and $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$8 is the optimal hedging strategy via the Galtchouk–Kunita–Watanabe projection. When the “pure-hedge condition” $$\frac{dX_t}{dt} = V(t, X_t),\qquad X_0\sim\gamma^{\mu,\Sigma},$$9 is satisfied, the martingale residual vanishes and hedging is risk-free (Okhrati et al., 2015).

5. Discrete-Time FS Decomposition, Stability, and Asymptotic Expansions

In discrete time, on a finite probability space with prices $$V(t,x) = \frac{x-\mu + \Sigma S(t,x)}{t},\quad S(t,x) = \nabla_x \ln \int_{\mathbb{R}^d}\varphi^{t y+(1-t)\mu,\, (1-t^2)\Sigma}(x) p(y)\, dy.$$0, the (discrete) FS decomposition is

$$V(t,x) = \frac{x-\mu + \Sigma S(t,x)}{t},\quad S(t,x) = \nabla_x \ln \int_{\mathbb{R}^d}\varphi^{t y+(1-t)\mu,\, (1-t^2)\Sigma}(x) p(y)\, dy.$$1

with orthogonality $$V(t,x) = \frac{x-\mu + \Sigma S(t,x)}{t},\quad S(t,x) = \nabla_x \ln \int_{\mathbb{R}^d}\varphi^{t y+(1-t)\mu,\, (1-t^2)\Sigma}(x) p(y)\, dy.$$2. The optimal hedge is determined via backward sequential regression:

$$V(t,x) = \frac{x-\mu + \Sigma S(t,x)}{t},\quad S(t,x) = \nabla_x \ln \int_{\mathbb{R}^d}\varphi^{t y+(1-t)\mu,\, (1-t^2)\Sigma}(x) p(y)\, dy.$$3

(Boese et al., 2020). This procedure recovers the classical delta hedge in binomial models and yields the least-squares optimal strategy in incomplete markets. The FS decomposition is stable under small perturbations of price dynamics, and explicit first-order corrections can be derived under parametric perturbations of drift and volatility.

6. Applications to Sampling and Machine Learning

Sampling via Föllmer flow enables efficient transformation of Gaussian samples to arbitrary target distributions. The deterministic nature of the flow permits fitting the map $$V(t,x) = \frac{x-\mu + \Sigma S(t,x)}{t},\quad S(t,x) = \nabla_x \ln \int_{\mathbb{R}^d}\varphi^{t y+(1-t)\mu,\, (1-t^2)\Sigma}(x) p(y)\, dy.$$4 with a deep neural network, yielding a one-step generator for direct sampling after supervised training. Föllmer flow can also be used as a warm-start for MCMC, providing initial states closely distributed according to complex multimodal targets, thus reducing mode collapse and improving mixing in Markov chain Monte Carlo algorithms (Ding et al., 2023).

7. Robustness, Change of Measure, and Effect of Model Parameters

The error bounds and convergence properties in FS-based hedging are robust under equivalent change of measure, provided the Radon–Nikodym derivative satisfies a reverse Hölder inequality. The effect of the Lévy measure, payoff regularity, jump intensity, and grid selection calibrate the theoretical and practical convergence rates. Fine tuning these parameters enables nearly optimal hedging performance in heavy-tailed jump models (Thuan, 2020).

---

Föllmer’s approach provides a unifying set of principles for constructing measure-valued flows for sampling, for local risk-minimization in incomplete markets via the Föllmer–Schweizer decomposition, and for controlled and provably convergent discretizations in both continuous and discrete finance models. Its effectiveness is enhanced through explicit representations, robust discretization analyses, and algorithmic compatibility with modern machine learning and Monte Carlo methods.