Full grid solution for multi-asset options pricing with tensor networks

This paper demonstrates that quantized tensor trains (QTT) overcome the curse of dimensionality in multi-asset option pricing by enabling full-grid solutions for the Black-Scholes PDE on up to 15 assets with high accuracy and polynomial scaling, offering a viable alternative to Monte Carlo methods.

The Gist

Imagine you are a financial architect trying to predict the future value of a complex building made of many different materials (stocks, bonds, currencies). In the world of finance, this is called pricing multi-asset options.

For decades, architects have struggled with a problem called the "Curse of Dimensionality." Here is the simple version of that curse:

• The Old Way (The Grid): Imagine you are trying to map a city. If you have one street, you need 100 points to map it. If you have two streets crossing, you need 100 x 100 = 10,000 points. If you have 5 streets, you need 100 x 100 x 100 x 100 x 100 = 10 billion points.
• The Problem: To get an accurate answer, you need to calculate the value at every single one of those points. With just 3 or 4 assets, your computer runs out of memory and time. It's like trying to count every grain of sand on a beach by picking them up one by one.
• The Current Workaround: Because the "grid" method is too slow, bankers usually use Monte Carlo simulations. This is like throwing darts at a board to guess the shape of the city. It's fast, but it's noisy, imprecise, and you only get a few data points. You can't see the whole picture, and if you want to know what happens if the wind blows slightly differently, you have to throw the darts all over again.

The New Solution: The "Quantum" Compression Trick

This paper introduces a new method using Tensor Networks (specifically called Quantized Tensor Trains or QTT). Think of this not as counting every grain of sand, but as realizing the beach has a hidden pattern.

Here is the analogy:

1. The "Zipper" Analogy (Compression)

Imagine you have a massive, high-resolution photo of a city. A normal computer tries to store every single pixel (Red, Green, Blue) for every inch. That's the "Full Grid." It's huge.

The QTT method is like a super-smart zipper. It realizes that the sky is mostly blue, the grass is mostly green, and the buildings have repeating patterns. Instead of storing every pixel, it stores the rules that generate the picture.

• The Magic: The more complex the city (the more assets you add), the more the "zipper" compresses the data. Instead of needing 10 billion points for 5 assets, this method might only need a few thousand "rules" to describe the whole thing perfectly.

2. The "Lego" Analogy (Building the Solution)

The authors built two different "construction kits" (solvers) to solve the math:

• Kit A: The Time-Stepper (The Staircase)
  • Imagine you are walking down a staircase from the future (when the option expires) back to today.
  • At each step, you calculate the value of the building based on the step below it.
  • Why it's good: It's very efficient for complex American options (where you can cash out early). It keeps the "Lego blocks" small and manageable at every step.
• Kit B: The Space-Time Solver (The Movie)
  • Instead of walking down the stairs one by one, imagine you have a time machine. You build the entire movie of the building's value from start to finish in one single shot.
  • Why it's good: It gives you the whole history at once. It's incredibly fast for simpler European options (where you can only cash out at the end).

Why This Matters (The "Aha!" Moment)

Before this paper, if you wanted to price a complex financial product involving 5 different stocks, you had to rely on the "dart-throwing" (Monte Carlo) method. You got a rough guess, and if you wanted to know how sensitive the price was to a tiny change in one stock (a "Greek"), you had to run the simulation again.

With this new method:

1. Full Picture: You get the entire map of the city, not just a few random points. You can see the value for any combination of stock prices instantly.
2. Instant Re-pricing: If the market moves 1% this morning, you don't need to re-run the whole calculation. You just "zoom in" on the map you already built. It's instant.
3. High Dimensions: The authors successfully ran this on a standard laptop for 3, 4, and even 5 assets. They estimate it could easily handle 10 to 15 assets with a bit more computing power.
4. No Noise: Unlike the dart-throwing method, this gives a mathematically exact answer (within a tiny margin of error), making it perfect for risk management.

The Bottom Line

This paper is like discovering a new lens for a camera.

• Old Lens: You could only take clear photos of small, simple objects (1-3 assets). Anything bigger was blurry or took forever to develop.
• New Lens (QTT): You can now take crystal-clear, high-definition photos of massive, complex scenes (5+ assets) in seconds, right on your laptop.

It turns a problem that was previously impossible to solve exactly into a routine calculation, allowing banks and traders to understand complex risks with a clarity they've never had before.

Technical Summary

1. Problem Statement

The pricing of multi-asset options (e.g., basket, max-min, or correlation-dependent options) using the Black–Scholes partial differential equation (PDE) is historically limited by the curse of dimensionality.

• Classical Limitations: Standard grid-based methods (finite differences, finite elements) require a grid size scaling as $O(N^d)$, where $N$ is the number of discretization points per dimension and $d$ is the number of underlying assets. This exponential growth restricts practical applications to $d \leq 3$.
• Current Industry Practice: For $d > 3$, practitioners rely on Monte Carlo (MC) methods. While MC handles high dimensions, it suffers from stochastic noise, slow convergence for tail events, and the inability to provide a full solution surface. MC yields prices only for specific initial states, requiring re-runs for new scenarios, and makes computing dense "Greeks" (sensitivities) difficult and noisy.
• Sparse Grid Limitations: Sparse grid techniques reduce complexity to $O(N(\log N)^{d-1})$ but do not provide a full grid representation, making tasks like interpolation for re-pricing or computing cross-sensitivities complex bottlenecks.

2. Methodology

The authors propose a deterministic approach using Quantized Tensor Trains (QTT) to represent the high-dimensional PDE, operators, and boundary conditions. This transforms the exponential storage and arithmetic complexity into a polylogarithmic dependence on the grid size.

Core Components:

• QTT Representation: The method reshapes the $d$-dimensional grid (size $2^c$) into a tensor with $d \times c$ binary modes. Functions and operators are decomposed into a chain of low-rank cores (Tensor Train format).
• Operator Construction: The Black–Scholes differential operator is discretized using implicit finite differences. Crucially, the resulting matrix (a sum of Kronecker products of tridiagonal Toeplitz matrices) admits an exact QTT representation with ranks that scale polynomially in $d$ (specifically $O(d^2)$) but are independent of the grid size $N$.
• Payoff and Boundary Conditions:
  • Smooth terms (exponentials) are constructed analytically with rank-1 QTT.
  • Non-linear terms (the $\max$ function in payoffs and early-exercise conditions) are handled using the TT-Cross algorithm, which samples a minimal number of points to construct a low-rank approximation without evaluating the full grid.
• Linear Solvers: The resulting linear systems are solved using the Alternating Linear Scheme (ALS) or its variant MALS (Multi-ALS), which optimizes tensor cores iteratively while keeping the rest fixed.

Two Solver Architectures:

1. Time-Stepping QTT Solver:
   • Solves the PDE sequentially in time (backward Euler).
   • At each step, it solves $A w_{i+1} = b_i$.
   • For American options, the early-exercise condition ($V \geq \text{Payoff}$) is enforced at every step via a TT-Cross projection.
   • Best for: $d > 3$, where the rank growth of the space-time formulation becomes prohibitive.
2. Space-Time QTT Solver:
   • Treats time as an additional spatial dimension, solving the entire space-time domain simultaneously in one shot.
   • Produces the full temporal evolution of the option price.
   • Best for: $d \leq 3$ and European options, offering superior speed and the full solution surface.

3. Key Contributions

• Breaking the Dimension Barrier: The first demonstration of full-grid solutions for multi-asset Black–Scholes PDEs with $d = 3, 4, 5$ (and theoretically up to 10–15) running on a standard personal computer (laptop).
• Full Grid & Dense Greeks: Unlike MC or sparse grids, this method provides the entire solution surface. This allows for:
  • Instant re-pricing of new spot prices via interpolation without re-running the solver.
  • Computation of dense, noise-free Greeks (Delta, Gamma, cross-sensitivities) across the entire grid.
• Rank Analysis: Proved that the QTT ranks for the Black–Scholes operator scale polynomially with $d$ ($O(d^2)$) and are independent of grid resolution. Numerical experiments show ranks remain low (e.g., $\chi \approx 10\text{--}12$) even for $d=5$.
• Algorithmic Framework: Developed two complementary solvers (Time-Stepping and Space-Time) and integrated the handling of non-linearities (American options) directly within the tensor framework.

4. Results

The authors benchmarked their solvers on a MacBook Pro (M3 Pro) against high-accuracy references (Gauss-Hermite quadrature for European, Longstaff-Schwartz MC for American).

• Accuracy: Achieved pricing errors of 1–2% for basket and max-min options in dimensions $d=3, 4, 5$.
• Performance:
  • European Options: The space-time solver was faster for $d \leq 3$. For $d=5$, the time-stepping solver solved the problem in ~433 seconds with 1% error.
  • American Options: The early-exercise condition added only a ~30% overhead to the runtime. The TT-Cross projection for the $\max$ function was efficient, keeping ranks controlled.
  • Scalability: A 5-asset problem with a grid of $2^{45}$ points (impossible for classical dense solvers due to RAM requirements of ~128 TB) was solved comfortably in memory.
• Greeks: Computing Greeks required negligible additional time (milliseconds) once the solution grid was built, as it involved applying low-rank derivative operators.

5. Significance and Impact

• Deterministic High-Dimensional Pricing: This work establishes QTT as a viable, deterministic alternative to Monte Carlo for high-dimensional derivatives, eliminating statistical noise and providing a complete view of the risk landscape.
• Operational Efficiency: The ability to build a full grid during market hours on commodity hardware enables instant re-pricing and real-time risk management (intraday Greeks), a significant improvement over overnight MC workflows.
• Calibration and Validation: The full-grid representation supports calibration to entire volatility/correlation surfaces rather than isolated points, improving stability in inverse problems.
• Future Extensions: The framework is generalizable to more complex market models, including local volatility, jump-diffusion (Merton), and SABR models, by adapting the discretization and operator splitting techniques.

In summary, the paper demonstrates that tensor network methods can overcome the curse of dimensionality in financial engineering, making full-grid, high-dimensional option pricing a practical reality on standard hardware.