TL;DR
This paper explores the application of quantum machine learning (QML) using a variational quantum classifier on IBM's quantum hardware and simulators for analyzing data from the Large Hadron Collider (LHC). The QML approach shows comparable performance to classical machine learning methods like SVM and BDT in distinguishing signal from background events in two key LHC analyses, demonstrating its potential for future high-energy physics research.
What Problem Does This Paper Address?
The Large Hadron Collider (LHC) experiments are constantly searching for new physics, which often involves identifying rare "signal" events amidst a massive "background" of known processes. Machine learning (ML) algorithms are crucial for this task, but the sheer volume and complexity of data can push the limits of classical computing. This paper investigates whether quantum machine learning, specifically a variational quantum classifier, can offer a viable alternative or complementary approach for these high-energy physics (HEP) analyses, leveraging the unique capabilities of quantum computers.
Methods
The core of this research is the quantum variational classifier (QVC) algorithm. The QVC works by:
- Encoding Data: Classical input data (features from LHC events) are encoded into a quantum state using a "feature map" circuit ($U(\mathbf{z})$). This circuit maps the classical data into a high-dimensional quantum feature space.
- Parameterized Quantum Circuit: A "variational circuit" ($W(\boldsymbol{\theta})$) with adjustable parameters ($\boldsymbol{\theta}$) is applied to this quantum state. This circuit is analogous to the "layers" or "weights" in a classical neural network.
- Measurement and Classification: The resulting quantum state is measured, and a specific operator is used to classify the event as either signal or background. The output is a binary label.
The key novelty lies in applying this QVC algorithm to real-world HEP datasets. The researchers used a 10-qubit setup on both IBM's quantum simulators and actual quantum hardware. To handle the input data, they employed Principal Component Analysis (PCA) to reduce the dimensionality of the kinematic variables to match the number of available qubits, followed by a MinMaxScaler to normalize the data. For training, they used a Spall's simultaneous perturbation stochastic approximation (SPSA) algorithm to optimize the variational circuit parameters.
Datasets
The paper uses simulated datasets from two specific LHC flagship physics analyses:
- $t\bar{t}H$ analysis: This involves the production of a Higgs boson in association with a top quark pair. The signal events are characterized by specific decay channels ($H \rightarrow \gamma\gamma$ and $t\bar{t}$ decays into jets).
- $H \rightarrow \mu^{+}\mu^{-}$ analysis: This focuses on Higgs boson decays into two muons, which probes its coupling to second-generation fermions.
For both analyses, the signal and background events were generated using Madgraph5_aMC@NLO and Pythia6, simulating proton-proton collisions at 13 TeV. The detector response was simulated using Delphes. The training and testing datasets each consisted of 100 events.
Key Insights and Ideas
- Quantum Advantage in Feature Space: The QVC leverages quantum circuits to explore an exponentially large feature space, which could potentially offer advantages over classical methods for complex data patterns.
- Hardware Noise is a Factor: While the QVC shows promise, the performance on actual quantum hardware is influenced by device noise. The researchers implemented error mitigation techniques to address this.
- Proof of Principle: The study serves as a crucial "proof of principle," demonstrating that QML can be applied to realistic HEP datasets and achieve meaningful results, even with limited qubit counts and small training sets.
- Comparable Performance to Classical ML: The most striking insight is that, with current technology and small datasets, the QVC performs on par with well-established classical ML algorithms like Support Vector Machines (SVM) and Boosted Decision Trees (BDT). This suggests QML is not yet outperforming classical methods but is a viable contender.
Key Results
- Simulator Performance: On the IBM quantum simulator with 10 qubits, the QVC achieved an Area Under the ROC Curve (AUC) of 0.81 ± 0.04 for the $t\bar{t}H$ analysis and 0.83 ± 0.05 for the $H \rightarrow \mu^{+}\mu^{-}$ analysis. These results were comparable to classical SVM (0.83 ± 0.04 and 0.82 ± 0.03, respectively) and BDT (0.83 ± 0.06 and 0.80 ± 0.06, respectively).
- Hardware Performance: On IBM's quantum hardware ("ibmq_boeblingen" and "ibmq-paris"), the QVC achieved AUCs of 0.82 for $t\bar{t}H$ and 0.81 for $H \rightarrow \mu^{+}\mu^{-}$. These results were in good agreement with the simulator's performance, indicating that the hardware, despite noise, could effectively learn to distinguish signal from background.
- Improved Signal Significance: Applying a selection based on the QVC discriminant improved the signal significance ($S/\sqrt{B}$) by approximately 50% for the $t\bar{t}H$ analysis and 75% for the $H \rightarrow \mu^{+}\mu^{-}$ analysis, compared to no selection.
- Training Convergence: The loss function evolution showed that the quantum algorithm, both on the simulator and hardware, was learning and converging, indicating it was effectively differentiating between signal and background events.
Limitations and Future Work
- Limited Qubit Count and Depth: The study was constrained to 10 qubits and limited circuit depths due to current hardware limitations and noise. This restricts the complexity of the models that can be implemented.
- Small Training Samples: The use of only 100 training events per dataset is a significant limitation. Larger datasets are crucial for robust ML model training.
- Hardware Noise: While error mitigation was applied, quantum hardware noise still impacts performance.
- Training Time: Training on quantum hardware was significantly slower than classical algorithms (200 hours for 500 iterations vs. much faster for classical methods).
The authors envision several future directions:
- More Qubits and Larger Datasets: Exploring QML with a larger number of qubits and more extensive datasets to potentially unlock "speed up" advantages and outperform classical classifiers.
- Quantum Kernel Methods: Investigating other QML algorithms, such as the quantum kernel classifier, which has shown promise.
- Mitigating Noise: Continued efforts to improve quantum hardware and develop more advanced error mitigation and correction techniques.
- Future LHC Applications: Applying QML to more complex future LHC analyses, such as measurements of Higgs boson self-couplings and searches for dark matter.
Summary updated