⚛️ quantum physics

Quantum Tensor Network Learning with DMRG

This paper introduces a global normalization condition to frame Matrix Product States as quantum states for machine learning and compares the effectiveness of gradient descent versus an adapted Density Matrix Renormalization Group (DMRG) algorithm for optimizing the resulting tensor network.

Original authors: Gustav J L Jäger, Martin B Plenio, Hans-Martin Rieser

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gustav J L Jäger, Martin B Plenio, Hans-Martin Rieser

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

In the quiet corners of physics where scientists study how particles behave together in vast groups, a powerful tool has emerged to make sense of the impossible. This tool, known as a tensor network, was originally designed to simulate the complex dance of quantum particles, helping researchers understand how matter settles into its most stable state. Think of it as a way to describe a massive, intricate system using a manageable set of building blocks, rather than trying to store every single detail in a computer's memory. Recently, scientists have begun to borrow these ideas for machine learning, hoping to teach computers to recognize patterns using the same efficient logic that governs the quantum world. The goal is to create learning systems that are not only powerful but also capable of running on future quantum computers, which operate on different principles than the devices we use today. However, a significant hurdle exists: quantum systems have a strict rule that the total probability of all possible outcomes must always add up to one, a condition that standard learning methods often ignore.

A team of researchers set out to bridge this gap by adapting a specific optimization technique called the Density Matrix Renormalization Group, or DMRG, for use in machine learning. This method, long a staple in quantum physics, is known for finding the best possible configuration of a system by adjusting its parts one by one. The researchers applied this approach to a neural network architecture based on matrix product states, a type of tensor network that arranges data in a chain-like structure. Their primary innovation was to enforce the strict normalization rule of quantum mechanics directly into the learning process. This meant that as the computer adjusted its internal parameters to learn from data, it was forced to keep the mathematical representation of the system perfectly balanced, just as nature demands. They tested two different ways to achieve this: a standard method that adjusts the system step-by-step using gradients, and their new modified version of the DMRG algorithm, which uses a more sophisticated mathematical shortcut to find the optimal solution.

To see if their ideas worked, the team trained these systems to recognize handwritten digits from the famous MNIST dataset, a standard benchmark where computers learn to identify numbers like zero through nine. They used a subset of 5,000 images, resizing them to a small grid of 49 pixels, and split them into training and testing groups. The results revealed a significant difference in how the systems behaved under quantum constraints. When they used a standard optimization method without the normalization constraint, the system achieved a test accuracy of 94.7 percent. However, the internal mathematical state of the system had grown so large and unbalanced that it no longer represented a valid quantum state. The paper notes that the norm of this state was approximately 3.9 million. When the researchers forced this unbalanced state to fit the rules by normalizing it, the resulting overlaps in the loss function became vanishingly small, leading to a trivial loss of approximately 0.5. This indicates that while the unconstrained method produced high accuracy metrics, the resulting model is not a valid quantum state and cannot be directly deployed on quantum hardware without losing its learned properties.

In contrast, the methods that enforced the normalization condition from the start behaved differently. Both the modified gradient descent and the new DMRG approach produced systems that respected the quantum rules, but they did not achieve the same high accuracy as the unconstrained method. The normalized gradient descent and the modified DMRG both reached a test accuracy of around 73 percent, with a loss value hovering near 0.36. While these numbers are lower than the unconstrained method, they represent a genuine, stable learning process that could actually run on a quantum computer. The researchers found that the modified DMRG algorithm, which uses a specific mathematical compression technique to solve the optimization problem, performed almost identically to the normalized gradient descent, confirming that the quantum-inspired shortcut was a viable alternative.

The study concludes that while forcing the system to obey quantum rules makes it less powerful in its current form compared to standard classical methods, it is a necessary step for the future. The high accuracy of the unconstrained method was achieved by ignoring the fundamental laws of the quantum world, resulting in a model that is not a valid quantum state. By accepting a lower accuracy now, the researchers have built a foundation that is mathematically sound and ready for quantum hardware. They acknowledge that more work is needed to improve performance and to develop even more complex ways to handle these constraints, but the path forward is clear: to bring machine learning to quantum computers, the algorithms must first learn to respect the strict rules of the quantum realm.

Technical Summary: Quantum Tensor Network Learning with DMRG

Problem Statement
Tensor Networks (TN), specifically Matrix Product States (MPS), have emerged as a promising machine learning architecture inspired by quantum many-body physics. While standard TN approaches often utilize gradient descent for optimization, they frequently lack the strict normalization conditions required for physical quantum states. In quantum mechanics, probability conservation dictates that quantum states and channels must be normalized. This paper addresses the challenge of adapting TN-based machine learning to satisfy these normalization constraints, thereby enabling the direct deployment of these models on quantum computers. The authors identify a gap in transferring local optimization techniques, specifically the Density Matrix Renormalization Group (DMRG) and its underlying Lanczos method, to the context of normalized quantum machine learning.

Methodology
The authors propose a framework where the MPS ansatz is constrained to yield a normalized vector upon contraction, effectively representing a quantum state. The core of their approach involves modifying standard optimization algorithms to respect the unit norm constraint ( $\langle \Psi | \Psi \rangle = 1$ ).

Loss Function and Constraints:
The authors define a mean square error loss function $L$ for classification tasks. Unlike unconstrained optimization where the optimum might be $A^{-1}|b\rangle$ , the normalization constraint requires the solution to lie on the unit sphere of the Hilbert space. The loss function is formulated as:
$L = \frac{1}{2} + \frac{1}{2N} \sum_{i=1}^N \left[ \langle \Psi | (|d_i\rangle\langle d_i|_D \otimes 1_B) | \Psi \rangle - 2\text{Re}(\langle \Psi |_{DB} (|d_i\rangle_D \otimes |l_i\rangle_B)) \right]$
where $|d_i\rangle$ represents input data and $|l_i\rangle$ represents labels.
Gradient Descent with Normalization:
The authors adapt the gradient descent algorithm to handle the normalization constraint. Instead of a simple update, they employ a projected gradient descent approach. After computing a gradient step $|n\rangle = |\Psi\rangle - \alpha |g\rangle$ , the state is renormalized: $|\Psi'\rangle = |n\rangle / \sqrt{\langle n|n\rangle}$ . To optimize efficiency, they derive a condition for the optimal step size $\alpha$ by setting the derivative of the loss with respect to $\alpha$ to zero, solving for $\alpha$ using Newton iterations.
Modified DMRG Algorithm:
Recognizing that Krylov subspace methods (like DMRG) are typically more effective than simple gradient descent, the authors develop a modified DMRG algorithm. Standard DMRG minimizes an energy expectation value but does not inherently account for the linear term in the machine learning loss function (Eq. 3).
- Post-Processing Step: After the Lanczos algorithm compresses the local effective Hamiltonian $A_s$ into a tridiagonal form $A'_s$ , the authors introduce a post-processing step. They solve for a scaling factor $x$ such that the gradient of the normalized state is perpendicular to the unit sphere.
- Analytical Solution: By decomposing the problem into the eigenbasis of the compressed matrix, they derive an equation for $x$ (Eq. 6) that ensures the normalization constraint is met. They prove that a unique global minimum exists for $x < 0$ and solve for it using Newton iteration, initialized carefully to ensure convergence.

Key Contributions

Normalization Condition: The introduction of a global normalization condition to the MPS ansatz, ensuring the resulting vector can be mapped to a valid quantum state.
Algorithmic Adaptation: The modification of both gradient descent and the DMRG/Lanczos algorithm to operate under this normalization constraint.
Comparative Analysis: A direct comparison between standard conjugate gradient descent, normalized gradient descent, and the proposed modified DMRG method.

Experimental Results
The methods were validated on a subset of the MNIST dataset (5,000 images, rescaled to 7x7), with 4,000 for training and 1,000 for testing. The models were optimized over three full two-site sweeps with a maximum bond dimension of 20.

Performance: The standard conjugate gradient descent achieved the highest accuracy (94.70% test accuracy) and lowest loss (0.08680). However, the resulting MPS had a norm of $\approx 3.9 \times 10^6$ . When normalized post-hoc, this led to vanishing overlaps and a trivial loss of $\approx 0.5$ , rendering the model ineffective for quantum deployment.
Normalized Methods: Both the normalized gradient descent and the modified Lanczos (DMRG) methods yielded identical training and test losses ( $\approx 0.36$ ) and accuracies ( $\approx 73\%$ ).
Observation: The introduction of the normalization condition significantly impacted the results, reducing accuracy compared to the unconstrained conjugate gradient method but ensuring the model remains physically valid (normalized).

Significance and Claims
The paper claims that implementing normalization conditions is a necessary step to make tensor network algorithms suitable for deployment on quantum computers. The authors position this work as a step toward adapting tensor network methods for more efficient quantum machine learning.

However, the authors are modest regarding the current performance. They explicitly state that the results show the performance of these normalized methods "needs to be increased to be competitive with other contemporary algorithms." They acknowledge that parity with classical tensor network methods (which do not enforce strict normalization in the same way) has not yet been achieved. The paper concludes that further research is required, including the study of more complex normalization conditions (such as those for quantum channels) and improvements in evaluation methods, to effectively bring machine learning to quantum computers.

Technical Summary: Quantum Tensor Network Learning with DMRG

More like this