Reusability Report: Optimizing T-count in General Quantum Circuits with AlphaTensor-Quantum

TL;DR

It is demonstrated that a general agent trained on circuits with varying qubit numbers outperforms agents trained on fixed qubit numbers, highlighting the method’s generalizability and its potential for broader quantum circuit optimization tasks.

Abstract

Quantum computing has the potential to solve problems that are intractable for classical computers, with possible applications in areas such as drug discovery and high-energy physics. However, the practical implementation of quantum computation is hindered by the complexity of executing quantum circuits on hardware. In particular, minimizing the number of T-gates is crucial for implementing efficient quantum algorithms. AlphaTensor-Quantum is a reinforcement learning-based method designed to optimize the T-count of quantum circuits by formulating the problem as a tensor decomposition task. While it has demonstrated superior performance over existing methods on benchmark quantum arithmetic circuits, its applicability has so far been restricted to specific circuit families, requiring separate, time-intensive training for each new application. This report reproduces some of the key results of the original work and extends AlphaTensor-Quantum's capabilities to simplify random quantum circuits with varying qubit counts, eliminating the need for retraining on new circuits. Our experiments show that a general agent trained on 5- to 8-qubit circuits achieves greater T-count reduction than previous methods for a large fraction of quantum circuits. Furthermore, we demonstrate that a general agent trained on circuits with varying qubit numbers outperforms agents trained on fixed qubit numbers, highlighting the method's generalizability and its potential for broader quantum circuit optimization tasks.

Figures (6)

• Figure 1: Reproducing AlphaTensor-Quantum. (a) The T-count reported in the original manuscript along with the results from experiments using the provided code. The training time to reach optimal performance on an NVIDIA A100 GPU is given in parentheses. Red numbers indicate where our experimental results do not match the originally reported values (see main text). (b) The evolution of the T-count during training. The light solid lines represent the reported result.
• Figure 1: Results for reproducing the AlphaTensor-Quantum by training one agent with the three circuits simultaneously. (a) The T-count reported in the paper and the results of the experiment with the provided code. The number in brackets indicates the total training time on a NVIDIA A100 GPU. The number in red shows the number where the reported value and the results of the experiments do not agree. (b) The evolution of the T-count during training. The light solid line shows the reported result.
• Figure 2: Average time for one step of AlphaTensor-Quantum training with gadgets on different GPU devices. Quadro RTX 6000 and Tesla V100 give an out-of-memory error for $15$ qubits. We compare with the baseline PyZX 2020_reducing and TODD heyfron_efficient_2018. Note that PyZX and TODD directly output the optimized circuit in the given time (e.g. around 0.06 seconds for 15 qubits). By contrast, AlphaTensor-Quantum requires a large number of training steps (e.g. between tens of thousands and several million in the original manuscript). Error bars, corresponding to one standard deviation across $10$ different circuits, are smaller than the marker size.
• Figure 2: Evaluation of single and general AlphaTensor-Quantum agents without gadgetization and three training types: only with synthetic demonstrations (Demo), only with reinforcement learning (RL), and both (Demo + RL). (a) The average T-count (lower is better) of the optimized quantum circuits in the evaluation set. The solid black line shows the average T-count of the baseline method PyZX 2020_reducing and TODD heyfron_efficient_2018. (b) The average T-count for each number of qubits. (c) The average improvement percentage (higher is better), which shows the percentage of circuits that have a strictly lower T-count when optimized with the agent compared to the baseline method. (d) The average improvement percentage for each number of qubits.
• Figure 3: Evaluation of single (random circuits, fixed qubit number) and general (random circuits, varying qubit number) AlphaTensor-Quantum agents with gadgetization and three training types: only with synthetic demonstrations (Demo), only with reinforcement learning (RL), and both (Demo + RL). (a) The average T-count (lower is better) of the optimized quantum circuits in the evaluation set. The solid black line shows the average T-count of the baseline method PyZX 2020_reducing and TODD heyfron_efficient_2018. (b) The average T-count for each number of qubits. (c) The average improvement percentage (higher is better), which shows the percentage of circuits that have a strictly lower T-count when optimized with the agent compared to the baseline method. (d) The average improvement percentage for each number of qubits. The error bar for (a) and (c) shows the 95% confidence intervals over different numbers of qubits and for (b) shows the 95% confidence intervals over $1000$ evaluation circuits.