Quantum RydKernel
Exponential concentration is a major obstacle overcome by the revolutionary quantum kernel method.
RydKernel, a novel quantum kernel method (QKM) developed by researchers at the Institut quantique, Sherbrooke, essentially addresses exponential concentration (EC), one of the biggest obstacles to the widespread use of quantum machine learning (QML). This innovation is a significant step towards achieving near-term quantum computers’ full capability in complex data analysis.
You can also read Dynamic Scheduling Boosts Hybrid quantum-classical computing
The Pervasive Problem of Exponential Concentration
In QML, kernel approaches have drawn a lot of interest as a potentially effective way to obtain a quantum advantage in data analysis. These techniques entail converting classical data into quantum states and calculating the “kernels,” or inner products, between them to gauge similarity. Finding the ideal parameters is guaranteed by the convex structure of the training landscape for kernel-based models, however this is predicated on the idea that kernel values can be effectively extracted from quantum hardware.
However, the phenomena of exponential concentration has seriously called into question this notion. Quantum kernel values spanning a range of input data can, in some cases, become exponentially concentrated (in terms of qubit count) towards a fixed value. This implies that as the number of qubits rises, off-diagonal kernel elements which are essential for differentiating between data points tend to disappear exponentially.
You can also read QCVV Quantum Characterization Verification And Validation
The consequences of EC are dire:
- Trivial Models: The statistical estimates of kernel values don’t reveal any significant information about the input data while training with a polynomial number of measurements, which is the realistic limit for modern devices. This results in a simple model in which the input data has no bearing on predictions made about invisible inputs.
- Poor Generalization: The trained model performs trivially on unknown data but successfully “hard-codes” training labels during optimization, leading to poor generalization capabilities. This obstacle must be overcome by exponentially increasing the number of measurement shots rather than by increasing the number of training data points.
- Analogy to Barren Plateaus: The issue of barren plateaus (BPs) in variational quantum algorithms (VQAs), where gradient magnitudes disappear exponentially, rendering training impossible, is comparable to EC in quantum kernels.
Four main sources have been shown to contribute to EC by research:
- Expressivity of Data Embedding: Excessively expressive embeddings result in roughly random and therefore orthogonal quantum states, which makes kernel values increasingly small.
- Global Measurements: Even with low expressivity and entanglement, fidelity kernels which depend on global measurements are vulnerable to EC.
- Entanglement: Concentration can also result from highly entangled encoded states, especially when paired with local quantum kernels such as projected quantum kernels.
- Noise: Particularly in circuits with polynomial depth, hardware noise can interfere with and destroy information, leading encoded states to concentrate towards the maximally mixed state.
You can also read Transformative Quantum Developments For Physics Sciences
RydKernel: A Solution Without Concentration
Ayana Sarkar, Martin Schnee, and Roya Radgohar, together with their collaborators, have created RydKernel, a QKM that is intrinsically free from exponential concentration but is nevertheless difficult to model classically, in response to these difficulties. Currently available analogue quantum kernels can directly apply this technique.
RydKernel takes advantage of the Rydberg blockade effect, which is a weak ergodicity-breaking many-body dynamics seen in coherently driven neutral atom arrays. The frequency shift (detuning) of a nearly-resonant driving laser applied to registers of highly interacting atoms is encoded with classical data. Usually, the system is started in a Néel state. The fidelity between two developed quantum states is measured by the kernel: Importantly, in order to avoid concentration effects that result from global observable measurements, the encoding time is set to be integer multiples of the system’s revival time.
You can also read A CMOS Silicon Spin Qubit Power Scalable Quantum Computers
Empirical and Analytical Validation
The researchers present strong proof of RydKernel’s lack of concentration:
- Analytical Toy Model: An approximate toy model of RydKernel’s dynamics shows that its variance is guaranteed to remain non-decreasing rather than decaying exponentially with growing system size.
- Numerical Simulations: These analytical predictions are validated by extensive numerical simulations. The absence of exponential concentration is demonstrated unequivocally by the RydKernel mean, which stays near 1, and its variance, which at most quadratically scales with system size.
- Practical Utility: Practical Utility When the encoding duration was long enough (T > 2.0 T_rev), RydKernel was able to effectively apply to the standard IRIS dataset classification job, attaining over 85% accuracy on both training and test datasets. This outcome confirms RydKernel’s ability to perform well on machine learning tasks, even though it does not yet surpass well-established classical approaches.
You can also read Projective Measurements Enable Topological quantum computing
Demonstrating Quantum Advantage
The classical simulability of any promising QML algorithm is a crucial component. Beyond moderately big systems, the researchers show that RydKernel is computationally impossible for classical computers to simulate.
- Highly entangled states arising from the Rydberg-blockaded dynamics are used in the fidelity computation.
- Even with sophisticated techniques like Matrix Product State (MPS)-based Time-Evolving Block Decimation (TEBD), simulating these dynamics becomes computationally prohibitive.
- At comparatively short durations, the Rydberg-blockaded dynamics, which originate from the |Z2> starting state, show volume-law entanglement. This necessitates a bond dimension that increases exponentially with the amount of qubits, quickly beyond the computational and memory capacities of classical systems.
- As an illustration of the possibility for a true quantum advantage, simulating a 45-qubit system the size of which is achievable with present neutral atom technology would require roughly one terabyte of memory and a significant amount of computational time. Moreover, it would be significantly more difficult to implement RydKernel in 2D for classical simulation.
You can also read Triple Quantum Dots: Superior Coherence And Qubit Control
Hardware Readiness for the Near Future
For use on current NAQCs, RydKernel is specially suited. These platforms have a reputation for:
- Connectivity that can be configured.
- Readout and high-fidelity quantum processes.
- Extended periods of decoherence.
A Loschmidt echo protocol can be used to implement the protocol. Preparing the |Z2> state (which can be accomplished using semi-local laser addressing or analogue techniques), embedding the first data point through a forward evolution, applying an approximate time-reversed evolution for the second data point (using a global Z gate), and then measuring in the computational basis are all steps in this process. The SWAP-test procedure is an approach that has been shown experimentally to reduce the necessity for time-reversed evolution.
Importantly, RydKernel’s experimental timings fall well within the present NAQC capabilities. With a preparation time of 500 ns and a global Z gate lifetime of 200 ns, a minimum implementation with two revivals for data encoding yields a total protocol duration of roughly 3660 ns. This easily falls within the 4500 ns coherence time (T2) that is typical for such devices. Furthermore, the Rydberg-blockaded dynamics are resistant to disorder and finite-temperature effects.
You can also read The Multi-QIDA: The Quantum Information Driven Ansatz
Outlook
By proving a QKM that tackles the crucial issue of exponential concentration while being experimentally practical and conventionally unsolvable, this study represents a major advancement. It emphasizes how crucial it is to use the special qualities of quantum many-body dynamics for QML. Further research will concentrate on expanding the approach to 2D neutral atom arrays, where classical simulations become much more difficult, and investigating the influence of quantum many-body scars and fragmentation in kernel performance.
The creation of RydKernel offers helpful advice for creating quantum embeddings that do not rely on concentration. It implies that the most effective way to achieve significant quantum advantage may be to use distinctly quantum approaches that make use of specialized quantum structures and symmetries rather than merely copying classical methods.
You can also read Quantum Time Transfer To Protect Against GNSS Problems
