Using non-convex optimization in quantum process tomography: Factored gradient descent is tough to beat

David A. Quiroga, Anastasios KyrillidisICRC 2023 · DOI

TL;DR

This paper introduces a novel non-convex optimization algorithm for quantum process tomography (QPT) that leverages matrix factorization. The proposed method, a factored gradient descent (FGD) variant, demonstrates faster convergence and higher fidelity compared to state-of-the-art convex optimization approaches, particularly in the presence of noise and with reduced measurement settings.

What Problem Does This Paper Address?

Characterizing the behavior of quantum computers is crucial for their development and reliable use, especially in the current NISQ (Noisy Intermediate-Scale Quantum) era. Quantum Process Tomography (QPT) is a key technique for this, aiming to understand how a quantum process affects quantum states. However, QPT is computationally demanding, requiring a large number of circuit configurations and measurements, which quickly becomes infeasible for systems with more than a few qubits. Existing methods often rely on convex optimization, which can be computationally expensive. This paper addresses the challenge of making QPT more efficient and scalable by proposing a non-convex optimization approach that exploits the inherent low-rank structure of many quantum processes.

Methods

The core of this work is a novel non-convex optimization algorithm for QPT based on the Burer-Monteiro (BM) factorization. This approach tackles the QPT problem by optimizing over the factored representation of the process matrix, $X = UU^\dagger$, where $U$ is a low-rank matrix. This factorization cleverly bypasses the need for explicit positive semi-definite constraints, which are computationally intensive in convex optimization.

The authors adapt the Projected Factored Gradient Descent (ProjFGD) method, originally developed for Quantum State Tomography (QST), to the QPT problem. The objective function is formulated to minimize the difference between measured and predicted measurement outcomes, incorporating a regularization term for the trace-preserving (TP) condition. The algorithm then iteratively updates the factors $U$ using a gradient descent step. Adaptive versions of gradient descent (adaGD) and factored gradient descent (adaFGD) are employed to dynamically adjust the step size, further improving efficiency.

The paper also investigates the noise resilience of their method by testing it against various noise models: depolarizing noise, Gaussian noise, coherent noise, and a general incoherent noise model.

Datasets

Not discussed in the paper. The authors generate synthetic data based on predefined quantum processes and noise models for their experiments.

Key Insights and Ideas

  • Factored Gradient Descent for QPT: Applying matrix factorization (Burer-Monteiro) to the process matrix in QPT is a novel approach that effectively turns a constrained convex optimization problem into an unconstrained non-convex one, leading to significant computational advantages.
  • Low-Rank Assumption as Noise Resilience: The explicit inclusion of a low-rank constraint via matrix factorization acts as a form of implicit regularization, making the algorithm more robust to noise compared to convex methods that don't leverage this prior knowledge.
  • Efficiency Gains with Reduced Measurements: The proposed FGD method shows superior performance even when using a significantly reduced number of measurements, aligning with compressed sensing principles and making QPT more practical for larger quantum systems.
  • FGD Outperforms GD in Noisy Scenarios: The experimental results consistently show that FGD achieves better fidelities and faster convergence than standard gradient descent (GD) across various noise models, especially in underdetermined settings.

Key Results

  • The proposed FGD algorithm achieves higher fidelities than state-of-the-art convex optimization approaches.
  • FGD demonstrates faster convergence, with improvements of up to 5x in some cases, particularly for depolarizing and Gaussian noise models.
  • The method is shown to be more noise-resilient, especially in the presence of depolarizing noise, where it converges to near-optimal solutions even with significant noise levels.
  • The algorithm performs well with a reduced set of initial states and measurement operators, requiring approximately $2.8^n$ circuit settings and $O(4^n)$ measurements for an underdetermined scenario, a significant reduction compared to traditional QPT.
  • While FGD shows inconsistent behavior with coherent noise, it still offers better estimates than GD in some cases.

Limitations and Future Work

  • Coherent Noise Inconsistency: The FGD algorithm exhibits unstable behavior with coherent noise, particularly in underdetermined settings, although it still manages to achieve better estimates than GD.
  • General Incoherent Noise: The algorithm struggled to accurately estimate quantum processes under a general incoherent noise model with large noise levels.
  • Future Work: The authors suggest providing theoretical guarantees for the FGD algorithm in QPT, conducting a complete study on measurement reduction using minimal initial state sets, and determining the absolute minimum number of measurement settings required for good fidelity in both determined and underdetermined cases. They also propose that using the depolarizing noise model might offer a good generalization for incoherent noise.

Summary updated

BibTeX
@inproceedings{quiroga2023using,
  author = {David A. Quiroga and Anastasios Kyrillidis},
  title = {{Using non-convex optimization in quantum process tomography: Factored gradient descent is tough to beat}},
  booktitle = {ICRC},
  year = {2023},
  doi = {10.1109/icrc60800.2023.10386455},
}
Figure