💰 quantitative finance

Dual Attainment in Multi-Period Multi-Asset Martingale Optimal Transport and Its Computation

This paper establishes the existence of dual optimizers for the multi-period, multi-asset martingale optimal transport problem under mild conditions, providing a rigorous theoretical foundation for robust financial pricing while demonstrating the practical solvability of large-scale instances via primal-dual linear programming.

Original authors: Charlie Che, Tongseok Lim, Yue Sun

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Charlie Che, Tongseok Lim, Yue Sun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to price a very complicated financial bet (a derivative) that depends on the future prices of several different stocks, not just one, and it depends on the entire path those stocks take over time, not just where they end up.

In the real world, we don't know the "true" future of the stock market. We only know the current prices of simple options (like standard bets on a single stock at a single date). These simple options give us "clues" about the possible future distributions of each stock individually, but they don't tell us how the stocks move together or how they behave over multiple days.

This paper tackles the problem of finding the safest possible price range for that complicated bet, using only the clues we have, without making up any extra rules about how the market behaves.

Here is the breakdown of what the authors did, using simple analogies:

1. The Puzzle: The "Martingale" Map

Think of the stock market as a hiker trying to cross a mountain range.

The Clues: We know exactly where the hiker starts (today's price) and we have a list of possible places they could be at specific checkpoints (tomorrow, next week, next month). These are the "marginal distributions."
The Rule: The hiker must follow the "Martingale" rule. In finance, this is like saying the hiker cannot have a secret advantage. On average, their next step must be exactly where they are standing right now. They can't drift systematically up or down; they just wander randomly.
The Goal: We want to know the highest and lowest possible price for a complex bet that pays out based on the hiker's entire journey.

2. The Problem: The "Dual" Solution

Mathematicians have a way to solve this by looking at it from two angles:

The Primal View: Trying to find the specific "worst-case" or "best-case" journey the hiker could take.
The Dual View: Trying to build a "safety net" using simple tools (like buying standard options and trading stocks dynamically) that guarantees you can cover the cost of the complex bet, no matter what path the hiker takes.

For a long time, mathematicians knew these two views should match (Duality). But they couldn't prove that a perfect "safety net" (a dual optimizer) actually exists for complex, multi-stock, multi-day scenarios. It was like knowing a perfect shield exists in theory, but being unable to actually build it. Without this proof, you can't be 100% sure your hedging strategy is mathematically sound.

3. The Breakthrough: Proving the Shield Exists

The authors of this paper proved that yes, you can always build this perfect safety net, provided the market clues aren't "broken" (a condition they call "irreducibility").

The Analogy: Imagine you are trying to build a fence around a wandering dog. You know where the dog might be at 1 PM, 2 PM, and 3 PM. You want to build a fence using only straight planks (simple options) and a moving gate (dynamic trading) that will catch the dog no matter which path it takes, as long as it follows the "no drift" rule.
The Result: They proved that for any number of dogs (stocks) and any number of time checkpoints, there is always a way to arrange your planks and gates to perfectly cover the dog's path. This is the Dual Attainment.

4. The Computer Test: Solving the Giant Puzzle

Proving it exists is one thing; actually finding the solution for a real-world problem is another. These problems are like trying to solve a Sudoku puzzle where the grid is the size of a city, and the rules change every second.

The Challenge: Traditional computers get overwhelmed by the sheer number of possibilities (the "curse of dimensionality").
The Solution: The authors used a super-fast, modern algorithm called PDLP (Primal-Dual Linear Programming) running on powerful graphics cards (GPUs).
The Test Case: They applied this to a real financial product called a "Worst-of Autocallable Option."
- What is it? A bet on two stocks (S&P 500 and NASDAQ). If the worse of the two stocks drops too low, you lose money. If it stays high, you get paid. But if it gets too high, the bet ends early.
- The Result: Their computer successfully calculated the exact price bounds and the specific "safety net" strategy (which options to buy and how to trade) for this complex product. The math worked perfectly, with the "gap" between the theoretical price and the calculated price being virtually zero.

Summary

In simple terms, this paper does two main things:

Mathematically: It proves that for complex, multi-stock financial bets, there is always a mathematically perfect way to construct a hedging strategy (a safety net) using standard market tools, as long as the market data is consistent.
Practically: It shows that we can actually compute these strategies for real-world, high-stakes products using modern supercomputers, confirming that the theory works in practice.

They didn't invent a new financial product or predict the future; they simply proved that the "mathematical safety net" for these complex bets is real, and showed us how to build it.

Technical Summary: Dual Attainment in Multi-Period Multi-Asset Martingale Optimal Transport and Its Computation

Problem Definition

The paper addresses the multimarginal, multi-asset Martingale Optimal Transport (MOT) problem, a fundamental challenge in the mathematical theory of model-independent pricing and hedging. Unlike classical Optimal Transport (OT), MOT requires the transport plan to satisfy the martingale property, reflecting risk-neutral valuation constraints.

The specific problem considers a financial market with $d$ underlying assets observed over $N$ time periods. The market reveals the marginal distributions of each asset at each maturity, but the joint distribution (the market model) remains unknown. The goal is to find the optimal transport plan $\pi$ that minimizes (or maximizes) the expected payoff of a path-dependent derivative $c(X)$ , subject to:

Marginal Constraints: The plan must match the observed marginal distributions $\mu_{t,i}$ for every asset $i$ and time $t$ .
Martingale Constraints: The conditional expectation of the next asset price must equal the current price ( $E[X_{t+1} | X_t] = X_t$ ).

A central theoretical question in this domain is dual attainment: the existence of an optimal dual pair $(\phi, h)$ representing semi-static hedging strategies (European options $\phi$ and dynamic trading strategies $h$ ) that achieve the dual bound. While duality (equality of primal and dual values) is well-established, proving the existence of dual optimizers in general multi-asset, multi-period settings has remained an open and challenging problem.

Methodology

Theoretical Framework

The authors establish dual attainment under mild regularity and irreducibility conditions. The proof strategy extends techniques from the single-asset, two-period case to arbitrary dimensions ( $d$ ) and time steps ( $N$ ).

Approximating Dual Maximizers: The authors construct a sequence of "nice" dual candidates $(\phi_n, h_n)$ where the dual values converge to the optimal value. These candidates satisfy the pathwise inequality required for sub-hedging.
Irreducibility and Convex Order: A key structural assumption is that the marginal distributions form an irreducible sequence in the convex order. This property ensures that the potential functions of the marginals behave regularly, preventing pathological cases where mass is "trapped" in disconnected components.
Local $L^1$ Bounds and Compactness: The core technical innovation involves deriving local $L^1$ bounds for the approximating potentials on compact intervals. By utilizing the irreducibility of the marginals, the authors prove that the sequence of dual potentials is locally bounded in $L^1$ .
Pointwise Convergence: Leveraging these bounds and Komlós's lemma, the authors extract a subsequence that converges pointwise almost everywhere to a limiting function. This limit serves as the dual optimizer.
Handling the Dual Class: The paper notes that the dual optimizer may not strictly belong to the original admissible class $\Psi$ (due to infinite-dimensional compactness issues), but the components are real-valued and measurable, sufficient for the financial interpretation of hedging strategies.

Numerical Implementation

To support the theoretical findings, the paper employs a computational approach using the Primal-Dual Linear Programming (PDLP) algorithm.

Discretization: The continuous MOT problem is discretized over a finite grid of asset prices and time steps, formulating it as a large-scale linear program (LP).
Relaxed Constraints: To handle discretization errors, the martingale constraints are relaxed with a tolerance parameter proportional to the grid size.
Solver: The authors utilize the PDLP solver (implemented in NVIDIA cuOpt), a first-order method designed for large-scale, sparse LPs. It uses adaptive step sizes, restarts, and diagonal preconditioning to handle the high-dimensional, path-dependent structure without explicit matrix factorization.

Key Contributions

Existence of Dual Optimizers: The primary contribution is a rigorous proof of dual attainment for the multimarginal, multi-asset MOT problem. This extends previous results, which were largely limited to single-asset or two-period settings, to arbitrary numbers of assets and time periods.
Theoretical Foundation for Robust Hedging: By establishing the existence of dual optimizers, the paper provides a mathematical guarantee that optimal sub- and super-hedging strategies exist for complex, path-dependent derivatives in high-dimensional settings.
Numerical Validation: The paper presents the first demonstration of applying the PDLP algorithm to large-scale multimarginal MOT problems. It validates the theoretical results by solving discrete instances with high precision.
Case Study on Worst-of Autocallables: The authors apply their framework to price a worst-of autocallable option referencing the S&P 500 and NASDAQ 100. This serves as a realistic test case for multi-asset, path-dependent products.

Results

Theoretical Results

Theorem 3.1 confirms that under irreducibility and lower semicontinuity of the cost function, a dual optimizer exists.
The dual optimizer $(\phi, h)$ satisfies the pathwise inequality for all price paths and achieves equality (replication) $\pi$ -almost surely for any primal optimizer $\pi$ .
The proof relies on the construction of approximating sequences and the extraction of limits via local compactness arguments derived from the irreducibility of marginals.

Numerical Results

Accuracy: The PDLP solver achieved near-machine precision. For the 2-asset, 3-time-step worst-of autocallable case, the duality gap was on the order of $10^{-13}$ to $10^{-14}$ .
Feasibility: Both primal and dual infeasibility norms were maintained well below $10^{-8}$ , confirming that the solver strictly enforced marginal and martingale constraints within the discretization tolerance.
Scalability: The use of GPU-accelerated PDLP allowed for the efficient solution of high-dimensional instances that would be infeasible on standard CPU-based solvers.
Case Study: The numerical experiment successfully computed tight price bounds for the worst-of autocallable option, demonstrating the practical viability of the theoretical framework.

Significance and Claims

The paper claims to advance the mathematical understanding of MOT by resolving the dual attainment problem in the general multi-asset, multi-period setting. This is significant because:

Robust Pricing: It provides a rigorous foundation for pricing complex derivatives without relying on specific parametric models, thereby capturing genuine model uncertainty.
Hedging Strategies: The existence of dual optimizers translates directly into the existence of explicit, model-independent hedging strategies using tradable instruments (options and dynamic trading).
Practical Applicability: The numerical experiments demonstrate that these theoretical advances are computationally tractable for real-world financial products, such as structured notes and autocallables, which depend on the joint evolution of multiple assets.

The authors maintain a modest stance regarding the computational aspect, noting that while they demonstrate the feasibility of solving these problems using PDLP, the primary focus remains on the theoretical development of dual attainment. The numerical work serves to corroborate the theoretical results rather than to propose new computational methodologies. Future work is suggested in refining theoretical bounds based on payoff structures and developing more efficient computational techniques for even larger scales.