⚡ electrical engineering

Tensor Train Quantum State Tomography using Compressed Sensing

This paper proposes a memory- and computationally efficient quantum state tomography method that utilizes low-rank block tensor train decomposition and compressed sensing to overcome the exponential scaling challenges of standard estimation techniques for a broad class of quantum states.

Original authors: Shakir Showkat Sofi, Charlotte Vermeylen, Lieven De Lathauwer

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shakir Showkat Sofi, Charlotte Vermeylen, Lieven De Lathauwer

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

The Big Problem: The "Infinite Library"

Imagine you have a quantum computer. To know how well it's working, you need to take a "snapshot" of its current state. In the quantum world, this snapshot is called a density matrix.

Think of this density matrix as a massive library of information.

For a small system (like 3 bits), the library has a few shelves.
But for a large system (like 12 bits), the number of books in that library doesn't just grow; it explodes. It grows exponentially.

Trying to read every single book in this library to figure out the state is impossible. It's like trying to count every grain of sand on a beach just to know how big the beach is. This is what scientists call the "curse of dimensionality." Standard methods try to read the whole library, which takes too much memory and time.

The Old Solutions: The "Shortcuts"

Scientists have tried to solve this by assuming the library isn't actually full of unique books. They assume the books follow a simple pattern (low-rank).

Method A (Convex Optimization): They try to find the pattern by checking every possible arrangement of books. It's accurate but incredibly slow, like trying to solve a 1,000-piece puzzle by trying every piece in every spot.
Method B (Factorization): They break the library down into smaller, manageable stacks. This is faster, but it's tricky to make sure the stacks still represent a valid quantum state (specifically, ensuring the "probability" never goes negative).

The New Solution: The "Block-TT" Train

The authors of this paper propose a new way to organize the library using a Tensor Train (TT) structure.

Imagine the massive library isn't one giant building, but a train made of many small, connected carriages.

The Train Cars (Tensor Train): Instead of storing the whole library in one place, the information is split across these cars. Each car only holds a small piece of the puzzle.
The Special Block: In this specific paper, they use a "Block-TT." Think of this as a train where one specific carriage is slightly different (a "block" carriage) that acts as a bridge.
The Magic Trick (Positive Semidefiniteness): In quantum mechanics, the "state" must be physically valid (probabilities can't be negative).
- Old methods often had to add extra rules or "brakes" to stop the math from breaking.
- This new method is like building the train out of a special material that cannot break. By constructing the state as a train connecting to its own mirror image (mathematically, $A \times A^H$ ), the result is guaranteed to be a valid, positive state automatically. You don't need to check the brakes; the train is built to be safe by design.

How It Works in Practice

The researchers tested this "Train" method against the old ways:

Speed: When they tried to measure the state of a large system, the old "Matrix" method took forever (exponential time). The new "Train" method was lightning fast (almost linear time). It's like switching from walking across the ocean to taking a high-speed bullet train.
Accuracy: Even with noisy data (like trying to hear a whisper in a loud room), the Train method reconstructed the quantum state just as well as, or better than, the other top methods.
Memory: Because the train only stores the small carriages and not the whole library, it uses a tiny fraction of the computer's memory.

The Bottom Line

The paper claims that by organizing quantum data into this specific "Block-TT Train" format, they can:

Save massive amounts of memory (no need to store the whole library).
Calculate much faster (no need to read every book).
Guarantee the result is physically valid without needing extra safety checks.

They tested this on simulated quantum systems with up to 12 qubits and showed that their method is a highly efficient, accurate way to "tomograph" (scan) quantum states, solving the problem of the "curse of dimensionality" for many types of quantum states.

Technical Summary: Tensor Train Quantum State Tomography using Compressed Sensing

Problem Statement
Quantum State Tomography (QST) is essential for benchmarking and verifying quantum devices by estimating the density matrix of a quantum system from measurement data. However, standard estimation methods face the "curse of dimensionality": as the number of qubits ( $N$ ) increases, the number of parameters in the density matrix grows exponentially ( $D = 2^N$ , requiring $O(D^2)$ parameters). While many quantum states of interest (e.g., ground states of local Hamiltonians, nearly pure states) possess low-rank structures that allow for efficient recovery with fewer measurements, existing scalable methods have limitations. Convex optimization approaches using nuclear norm penalization become computationally prohibitive for large systems. Factorization methods (e.g., Burer-Monteiro) improve efficiency but often struggle to enforce necessary physical constraints (Hermiticity, positive semidefiniteness, and unit trace) without complex projections. Furthermore, while tensor network representations like Matrix Product States (MPS) are efficient for pure states, extending them to mixed states while maintaining these constraints remains challenging.

Methodology
The authors propose a novel QST framework that parameterizes the density matrix $\rho$ using a Block Tensor Train (Block-TT) decomposition. This approach combines compressed sensing principles with tensor network algebra.

Parametrization: Instead of optimizing over the full density matrix, the method represents $\rho$ as the contraction of a Block-TT network $A$ with its Hermitian transpose $A^H$ (i.e., $\hat{\rho}_{TT} = A \bullet_n A^H$ ).
- The network $A$ is an $(N+1)$ -th order tensor where most cores are third-order, except for one fourth-order core at a specific position $n$ carrying an auxiliary index $K$ .
- This construction inherently guarantees that the resulting density matrix is positive semidefinite (PSD) and Hermitian, analogous to a Cholesky decomposition for matrices, without requiring explicit constraints during optimization.
- The unit trace constraint is managed via a Frobenius norm constraint on the factor $A$ ( $\|A\|_F \leq 1$ ).
Optimization: The QST problem is reformulated as a non-convex optimization problem minimizing the least-squares error between observed measurement outcomes and the model predictions.
- Objective: $\min_A \frac{1}{2} \|y - \mathcal{M}(\hat{\rho}_{TT})\|_2^2$ subject to $\|A\|_F \leq 1$ .
- Algorithm: A simple Gradient Descent (GD) algorithm is employed. In each iteration, the gradient is computed using tensor network contractions, and the updated factor is projected back onto the Block-TT format using a truncated SVD (TT-SVD) to maintain orthogonality and control ranks. A momentum factor is used to improve convergence.
Efficiency: The method leverages the TT algebra to compute expectation values ( $\text{Tr}(\rho E_m)$ ) efficiently. The computational complexity for evaluating measurements scales as $O((4R^2 + 2R^3)(\log_2 D + K))$ , where $R$ is the TT-rank. This is significantly more efficient than the $O(D^2)$ complexity of standard matrix trace computations for large $N$ , provided the state has a low TT-rank.

Key Contributions

Block-TT Parametrization: Introduction of a Block-TT representation for density matrices that inherently preserves positive semidefiniteness, eliminating the need for complex projection steps required by other factorization methods.
Scalability: Demonstration that the approach mitigates the curse of dimensionality for a broad class of quantum states (pure, nearly pure, and ground states) by reducing the parameter count from exponential to logarithmic scaling relative to the system size.
Algorithmic Framework: Development of a gradient-based optimization algorithm tailored for Block-TT networks, utilizing TT-SVD for rank truncation and normalization.
Comparative Analysis: A comprehensive comparison with state-of-the-art methods, including Convex Optimization (CVX), Maximum Likelihood Estimation (MLE), and Low-Rank Projected Gradient (LR) methods.

Results
Numerical experiments were conducted on systems ranging from 3 to 12 qubits with random Pauli measurements and Gaussian noise (SNR = 60 dB).

Accuracy: The proposed Block-TT method achieved competitive accuracy in recovering mixed states. In terms of Fidelity and Trace Distance, it outperformed MLE and showed overall similar performance to the Low-Rank (LR) projected gradient method.
Sampling Efficiency: At low sampling ratios ( $M/D^2 < 0.3$ ), the convex optimization (CVX) method performed best, but the Block-TT method remained competitive.
Computational Efficiency: The most significant finding was in computational time. While matrix-based measurement evaluation scales exponentially with the number of qubits, the Block-TT approach scales close to linearly. This allows the method to handle large-scale systems (e.g., 12 qubits) where matrix-based methods become computationally intractable.

Significance and Claims
The paper claims that the proposed Block-TT QST approach offers a practical solution for large-scale quantum state estimation by balancing memory efficiency, computational speed, and physical validity.

Modest Claims: The authors explicitly state that the method is efficient specifically for quantum states that can be well-approximated by low-rank tensor networks. They acknowledge that the current algorithm is a simple gradient descent and that more advanced optimization over TT manifolds could further improve performance.
Future Directions: The paper suggests that future work could include rigorous theoretical analysis of recovery guarantees and the development of structured tensor completion approaches for deterministic recovery.
Conclusion: The work demonstrates that parameterizing density matrices via Block-TT networks is a viable, memory- and computationally efficient alternative to traditional tomography, capable of recovering accurate mixed states without imposing additional constraints beyond the inherent structure of the decomposition.