An iterated Azéma-Yor type embedding for finitely many marginals

Jan Obłój, Peter SpoidaOxford 2017

Abstract

We solve the $n$-marginal Skorokhod embedding problem for a continuous local martingale and a sequence of probability measures $\mu_1,...,\mu_n$ which are in convex order and satisfy an additional technical assumption. Our construction is explicit and is a multiple marginal generalisation of the Azema and Yor (1979) solution. In particular, we recover the stopping boundaries obtained by Brown et al. (2001) and Madan and Yor (2002). Our technical assumption is necessary for the explicit embedding, as demonstrated with a counterexample. We discuss extensions to the general case giving details when $n=3$. In our analysis we compute the law of the maximum at each of the $n$ stopping times. This is used in Henry-Labordere et al. (2013) to show that the construction maximises the distribution of the maximum among all solutions to the $n$-marginal Skorokhod embedding problem. The result has direct implications for robust pricing and hedging of Lookback options.

Introduction

The Skorokhod Embedding Problem (SEP) is a foundational challenge in probability theory: given a stochastic process (such as Brownian motion) and a target probability distribution $\mu$, can one identify a stopping time $\tau$ such that the process stopped at $\tau$ follows distribution $\mu$?

While the single-marginal case ($n=1$) has been well-understood for decades, modern quantitative finance requires matching multiple distributions (marginals) at different time points—for instance, when market prices for call options are known at several maturities. This paper provides the first explicit, iterative solution to the $n$-marginal SEP. By generalizing the classic Azéma-Yor construction, the authors identify a sequence of stopping times $\tau_1 \leq \tau_2 \leq \dots \leq \tau_n$ that match a sequence of distributions $\mu_1, \dots, \mu_n$. This construction is specifically designed to maximize the distribution of the process's maximum, which is critical for determining model-independent price bounds for path-dependent financial instruments like Lookback options.

Methodology: Iterative Stopping Boundaries

The authors develop an iterative construction based on the spirit of the Azéma-Yor (1979) solution. For each marginal $i$, they define a stopping boundary function $\xi_i$. The process is stopped at time $\tau_i$ the first time it hits this boundary relative to its current maximum.

Visualization of the stopping boundaries $\xi_1, \xi_2, \xi_3$ for the 3-marginal case.

Unlike a simple repetition of the 1-marginal case, the $n$-th boundary $\xi_n$ is found by solving an optimization problem that incorporates all previous boundaries ($\xi_1, \dots, \xi_{n-1}$) and the associated "penalty" terms. To maintain an explicit and "clean" solution, the authors introduce a technical assumption (Assumption $\circledast$) on the measures. This ensures that the boundaries remain continuous and non-decreasing, preserving the strong relation between the process value and its maximum across all steps.

Theoretical Insights

A key insight of this work is that a "greedy" iteration of 1-marginal solutions is insufficient for global optimality. The $n$-th stopping time must account for the constraints imposed by all preceding marginals. The authors introduce the concept of a "binding constraint" index $\jmath_n(y)$, which identifies which previous marginal restricts the choice of the next stopping boundary at a given level of the maximum.

Furthermore, the paper derives an Ordinary Differential Equation (ODE) describing the distribution of the process's maximum at each stopping time. This ODE serves as the essential link between the geometric stopping boundaries and the target probability distributions.

Key Results

The main result (Theorem 2.6) formally proves that the constructed sequence $(\tau_i)$ embeds the measures $\mu_i$ and that the resulting process is a uniformly integrable martingale. This construction successfully recovers and unifies known results for $n=1$ (Azéma-Yor) and $n=2$ (Brown et al., 2001).

Additionally, the authors provide a counterexample involving three measures that violate Assumption $\circledast$, demonstrating that this technical condition is necessary for the explicit formula to hold. In cases where the assumption is violated, the interdependencies between marginals become significantly more complex, requiring a more technical, implicit approach.

Conclusion and Future Work

While the proposed construction is highly effective under Assumption $\circledast$, extending the explicit solution to more complex interdependencies remains a challenge. For $n > 3$, the problem becomes increasingly technical without this assumption. Future research may focus on developing computationally efficient methods for these edge cases or exploring iterated embeddings under alternative optimality criteria, such as variance minimization.

Summary updated

Abstract

We consider here an n-marginal Skorokhod embedding problem (SEP). We construct an explicit solution which has desirable optimal properties. The classical (one-marginal) SEP consists in finding a stopping time τ such that a given stochastic process (Xt) stopped at τ has a given distribution µ. For the solution to be useful (and non-trivial) one further requires τ to be minimal (cf. Ob l´oj (21, Sec. 8)). When X is a continuous local martingale and µ is centred in X0, this is equivalent to (Xt∧τ : t ≥ 0) being a uniformly integrable martingale. The problem dates back to the original work in Skorokhod (25) and has remained an active field of research since. New solutions often either considered new classes of processes X or focused on finding stopping times τ with additional optimal properties. This paper contributes to the latter category. We are motivated, as was the case for several earlier works in the field, by questions arising in mathematical finance which we highlight below.
BibTeX
@inproceedings{oboj2017iterated,
  author = {Jan Obłój and Peter Spoida},
  title = {{An iterated Azéma-Yor type embedding for finitely many marginals}},
  booktitle = {Oxford},
  year = {2017},
}
Figure