⚛️ quantum physics

Learning PDEs for Portfolio Optimization with Quantum Physics-Informed Neural Networks

This paper introduces a Quantum Physics-Informed Neural Network (QPINN) utilizing parameterized quantum circuits with tensor rank decomposition to solve the Merton portfolio optimization PDE, demonstrating that the quantum approach achieves higher accuracy and faster convergence with 80 times fewer parameters than classical PINN counterparts.

Original authors: Letao Wang, Abdel Lisser, Sreejith Sreekumar, Zeno Toffano

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Letao Wang, Abdel Lisser, Sreejith Sreekumar, Zeno Toffano

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 an investor trying to decide how to split your money between a safe savings account and a risky stock market. You want to maximize your future wealth, but the market is chaotic and unpredictable. In the world of finance, this is a famous puzzle called the Merton Portfolio Optimization Problem.

To solve it, mathematicians use a complex set of rules called Partial Differential Equations (PDEs). Think of these equations as a giant, multi-dimensional maze. Finding the "exit" (the perfect investment strategy) is incredibly hard.

Here is the story of how this paper proposes a new, super-efficient way to solve that maze using Quantum Physics and Neural Networks.

1. The Old Way: The Slow, Heavy Truck

Traditionally, computers try to solve these financial mazes by breaking them down into tiny grid squares (like a chessboard) and calculating the answer for every single square.

The Problem: As the problem gets more complex, the grid gets huge. The computer has to drive a heavy truck through every single square. It's slow, expensive, and often gets stuck in traffic (slow convergence).
The Alternative (Classical AI): Scientists tried using Physics-Informed Neural Networks (PINNs). Imagine a student trying to learn the maze by guessing and checking. They are faster than the truck, but they still struggle to learn the "shape" of the maze quickly, often needing millions of guesses to get it right.

2. The New Idea: The Quantum Magic Wand

The authors of this paper asked: "What if we could use the weird, powerful rules of quantum physics to solve this?"

They built a new tool called a Quantum Physics-Informed Neural Network (QPINN). But there was a catch: Quantum computers are currently small and fragile. Trying to run a massive calculation on them is like trying to carry a watermelon on a bicycle; it's too heavy and will break the bike.

3. The Secret Sauce: "Tensor Decomposition" (The Lego Trick)

This is the paper's biggest breakthrough. They realized that many financial solutions (like the Merton problem) have a special structure. They aren't just random chaos; they are built from independent parts that multiply together.

The Analogy:
Imagine you need to build a massive castle.

The Old Quantum Way: You try to build the whole castle as one giant, solid block of stone. It's too heavy to lift.
The Paper's New Way: They realized the castle is actually just a collection of separate Lego towers (walls, turrets, gates) that you can snap together.
- They call this Tensor Decomposition. Instead of building the whole complex solution at once, they build small, simple "univariate" (one-variable) Lego pieces and snap them together.

By using this "Lego" approach, they reduced the quantum resources needed from exponential (impossible) to polynomial (manageable). It's like realizing you don't need a crane to move the whole castle; you just need a small truck to move the Lego pieces.

4. The Two Models: The Quantum Robot and the Quantum Ghost

The team created two versions of their solver:

The Quantum-Inspired PINN (The Ghost):
- This model uses the "Lego" structure but runs on a classical computer (like your laptop).
- It's called "Quantum-Inspired" because it uses the logic of quantum math but doesn't need a quantum computer.
- Result: It solved the problem 80 times faster and with 80 times fewer parameters (less memory) than a standard AI, because it knew exactly how to build the Lego castle.
The QPINN (The Robot):
- This is the real deal. It runs on a quantum circuit.
- It adds a special "entanglement" layer (a quantum magic trick where parts of the system talk to each other instantly).
- Result: It was even better than the Ghost. It found the solution with higher accuracy and converged (learned) faster.

5. The Experiment: The Race

They tested these models on the Merton investment problem.

The Competitor: A standard, heavy-duty AI (Fully Connected PINN) with 481 parameters. It was like a giant, clumsy robot trying to learn the maze.
The Winners: Their Quantum models had only 7 parameters.
The Outcome: The tiny Quantum models (with just 7 parameters) beat the giant robot (with 481 parameters) every time. They learned the maze shape almost instantly and found the perfect investment strategy.

Why This Matters

Efficiency: You don't need a super-computer to solve complex financial problems anymore. A tiny, efficient model can do it better.
Future-Proof: This method is designed to work on the small, imperfect quantum computers we have today (and in the near future), not just the theoretical ones of the distant future.
The Lesson: Sometimes, the best way to solve a complex problem isn't to throw more power at it, but to understand its hidden structure (the "Lego" nature) and build a tool that fits that structure perfectly.

In a nutshell: The authors took a difficult financial math problem, realized it was made of simple building blocks, and built a tiny, super-efficient quantum-powered robot that could snap those blocks together faster and better than any giant classical computer could.

1. Problem Statement

The paper addresses the challenge of solving Partial Differential Equations (PDEs) in financial mathematics, specifically the Merton Portfolio Optimization problem.

Context: The Merton problem seeks an optimal investment strategy to maximize the expected utility of terminal wealth under stochastic market dynamics. This is mathematically transformed into a non-linear Hamilton-Jacobi-Bellman (HJB) PDE.
Limitations of Classical Methods:
- Numerical Solvers (FDM/FEM): Require high computational costs for high accuracy and struggle with the "curse of dimensionality."
- Classical PINNs (Physics-Informed Neural Networks): While mesh-free, they often suffer from slow convergence, difficulty capturing high-frequency features, and scalability issues in high dimensions. Furthermore, standard Fully Connected (FC) PINNs require massive parameter counts to achieve comparable accuracy.
Limitations of Existing Quantum Approaches:
- Fully Quantum Algorithms: Require fault-tolerant hardware and qRAM, which are not yet available.
- Current QPINNs: Often rely on Fourier series or hardware-efficient ansatzes (HEA) with hidden constraints. They lack theoretical guarantees that their hypothesis space contains an approximation of the target function. Additionally, implementing general multivariate polynomials on quantum circuits typically incurs exponential resource complexity (depth and parameters), making them infeasible for near-term devices.

2. Methodology

The authors propose a framework combining Quantum Machine Learning (QML) with Tensor Rank Decomposition to create efficient solvers for PDEs.

A. Mathematical Foundation: Tensor-Decomposed Polynomials

The core innovation is restricting the hypothesis space to Tensor-Decomposed (TD) Polynomials.

Definition: A multivariate polynomial $p(x)$ is expressed as a sum of products of univariate polynomials:
$p(x) = \sum_{r=1}^{R} \lambda_r \prod_{j=1}^{D} p_{r,j}(x_j)$
where $R$ is the tensor rank.
Motivation: Many PDE solutions (including the Merton HJB equation) possess separable structures that can be well-approximated by low-rank tensor decompositions.
Resource Complexity: While general multivariate polynomials require exponential circuit depth, the authors prove that TD polynomials can be implemented with polynomial complexity in terms of circuit depth and parameters, provided the tensor rank $R$ is moderate.

B. Quantum Architecture

The authors design two specific models:

Quantum Physics-Informed Neural Network (QPINN):
- Structure: Based on a Parameterized Quantum Circuit (PQC) implementing the TD polynomial.
- Enhancement: An entangling layer (controlled unitary $e^{-iH(\lambda)}$ ) is added to the circuit. This layer introduces non-separable multi-qubit correlations, expanding the hypothesis space beyond simple tensor products and increasing expressivity.
- Hypothesis Space: Contains all TD polynomials of a fixed degree and rank, plus additional entangled states.
Quantum-Inspired PINN:
- Structure: A classical simulation of the QPINN where the entangling layer is removed (or set to identity).
- Function: It strictly implements the TD polynomial structure.
- Significance: It serves as a classical baseline that shares the exact same "inductive bias" (structural assumption) as the quantum model but is computationally cheap to run.

C. Theoretical Guarantees

Existence of Approximation: The authors prove that the hypothesis space of their QPINN explicitly contains the set of all TD polynomials. Since many PDE solutions admit such a decomposition, the existence of a good approximation is guaranteed (unlike standard HEA-based QPINNs).
Complexity Analysis:
- Univariate: $O(L)$ depth.
- General Multivariate: Exponential depth $O(L^D)$ .
- TD Polynomial: Polynomial depth $O(RDL \log R)$ and parameters $O(RDL)$.

3. Key Contributions

Efficient Quantum Implementation: The paper provides a detailed resource analysis showing that implementing tensor-decomposed polynomials reduces quantum resource complexity from exponential to polynomial, making it feasible for Near-Term Intermediate-Scale Quantum (NISQ) devices.
Theoretical Guarantees: Unlike previous QPINN works, this model guarantees that the hypothesis space contains a suitable approximation for PDEs with separable solutions, based on the properties of tensor rank decomposition.
Dual Model Framework: The introduction of both a QPINN (with entanglement) and a Quantum-Inspired PINN (classically simulable) allows for a rigorous dissection of "quantum advantage" vs. "inductive bias advantage."
Application to Finance: The framework is successfully applied to the Merton portfolio optimization problem, a critical real-world financial PDE.

4. Experimental Results

The models were tested on the Merton HJB PDE with parameters: $r=0.02, T=1.0, \gamma=0.95, \mu=0.0219, \sigma=0.2$ .

Comparators:
- QPINN: 7 trainable parameters (6 from TD structure + 1 entangling).
- Quantum-Inspired PINN: 6 trainable parameters (TD structure only).
- Counterpart PINN: A classical PINN with the same TD inductive bias (6 parameters).
- FC PINN: A standard Fully Connected PINN with 5 hidden layers (481 parameters, ~80x more than the quantum models).
Performance:
- Convergence: Both QPINN and Quantum-Inspired PINN converged faster and achieved lower loss than the FC PINN, despite having 80 times fewer parameters.
- Accuracy: The QPINN achieved the highest accuracy, outperforming the Quantum-Inspired PINN. This demonstrates the benefit of the entanglement layer (quantum expressivity) in capturing complex features near boundaries.
- Inductive Bias: The Quantum-Inspired PINN outperformed the FC PINN, proving that the tensor-decomposed structure (inductive bias) is the primary driver of efficiency, not just the quantum hardware.
- Visuals: The QPINN surface closely matched the analytical solution across the entire domain, whereas the FC PINN showed significant deviations.

5. Significance and Conclusion

Resource Efficiency: The work demonstrates a viable pathway to solving structured PDEs on near-term quantum hardware by leveraging tensor decomposition to avoid exponential resource scaling.
Inductive Bias vs. Quantum Advantage: The study clarifies that for structured problems like portfolio optimization, the structural inductive bias (tensor decomposition) is crucial. The quantum advantage (entanglement) provides a further boost in expressivity and accuracy, allowing the model to outperform even large classical networks with the same structural bias.
Practicality: The "Quantum-Inspired" model suggests that even without quantum hardware, adopting these specific architectural biases can yield superior classical solvers. However, the QPINN shows that quantum entanglement can push performance beyond what is achievable classically with the same parameter count.
Future Work: The authors note that while expressivity is improved, the trainability (avoiding barren plateaus) and rigorous error bounds remain open areas for future research.

In summary, this paper presents a theoretically grounded and experimentally validated approach to solving financial PDEs, bridging the gap between quantum computing potential and practical near-term application through tensor-decomposed architectures.