Quantum Embeddings: Theory & Applications

• Quantum embeddings are nonlinear mappings from classical data to quantum state space using trainable circuits for enhanced feature representation.
• They employ specific circuit layers such as Ry rotations and entangling gates to compute quantum kernels and facilitate quantum metric learning.
• Empirical results show that QES-optimized architectures achieve competitive accuracy with fewer parameters and shallower circuit depths compared to classical models.

A quantum embedding is a nonlinear (often trainable) map from classical input space into a quantum (Hilbert) space, realized via a parameterized quantum circuit, with the goal of exploiting the representational and geometric power of quantum states for use in supervised, unsupervised, or downstream quantum/classical algorithms. Quantum embeddings serve as the foundational feature map in quantum machine learning, quantum representation learning, and quantum embedding theories for many-body physics, and appear in quantum-inspired classical models as efficient or interpretable ways of encoding structure.

1. Mathematical Formalism and Circuit Structure

Let $x\in\mathbb{R}^p$ denote a classical feature vector. A quantum embedding is a map

$$\phi: x \mapsto |\psi(x)\rangle \in \mathcal{H},\quad \mathcal{H} \cong \mathbb{C}^{2^n}$$

where $\mathcal{H}$ is the Hilbert space of $n$ qubits. Typically, one prepares $|\psi(x)\rangle$ via a parameterized quantum circuit (ansatz):

$$|\psi(x)\rangle = U(\theta, x)\,|0\rangle^{\otimes n}$$

with $U(\theta,x)$ combining input-dependent rotations, entangling unitaries, and potentially trainable blocks:

$$U(\theta, x) = U_w(\theta) \cdot U_E(\alpha) \cdot U_x(x)$$

• $U_x(x) = \bigotimes_{j=1}^{n} R_y(x_j)$ is the input feature block (typically $y$-axis rotations),
• $U_E(\alpha)$ is a sequence of CNOTs or other entangling gates,
• $U_w(\theta) = \bigotimes_{j=1}^{n} R_y(w_j)$ represents trainable single-qubit rotations.

The quantum kernel associated with $\phi$ is

$K(x, x') = |\langle\psi(x)|\psi(x')\rangle|^2,$

enabling kernel-based learning and quantum metric learning paradigms (Nguyen et al., 2021).

2. Search Space, Entanglement, and the QES Algorithm

The expressivity and efficiency of quantum embeddings are intimately tied to the structure of the entangling layers. Each entangling block $U_E(\alpha)$ can be represented as an ordered, directed multi-graph $G=(V,E)$ where $V$ are qubits and $E$ denotes the directed CNOTs. The "genotype" $\alpha$ is a length-$k$ sequence of directed edges (CNOTs), with $k$ termed the entanglement level.

For $N$ qubits, the total number of possible distinct directed CNOT edges is $E_\mathrm{tot}=N(N-1)$. The reduced search space of depth-$k$ embedding circuits (fixing $k$ entanglers per layer) is

$$|\Omega_\mathrm{reduced}(k)| = \binom{E_\mathrm{tot}}{k}\,k! = E_\mathrm{tot}!/(E_\mathrm{tot}-k)!$$

Searching this space is intractable for large $N$ or $k$.

Quantum Embedding Search (QES) tackles this via a sequential model-based optimization (SMBO) algorithm using a surrogate (e.g., TPE), which proposes candidate entanglement patterns maximizing an acquisition function (expected improvement), thereby efficiently exploring $|\Omega_\mathrm{reduced}(k)|$ for optimal embedding architectures (Nguyen et al., 2021).

3. Empirical Performance and Practitioner Guidelines

Extensive empirical benchmarking demonstrates QES-found architectures achieve or surpass hand-designed embeddings and approach classical model accuracy on synthetic, UCI, and hybrid autoencoder-low-qubit datasets:

• On a 4-dimensional synthetic dataset: QES-TPE (depth=2, $k=8$) achieved 83.5% accuracy, matching SVM/XGBoost and outperforming manual entangling designs by 2-7%.
• On the Iris dataset (4 dim, 3 class): QES-TPE (depth=2, $k=4$, 23 params) reached 95.33% accuracy, using $\sim2.5\times$ fewer parameters than the best fair NN baseline.
• On higher-dimensional tabular data autoencoded to 4 qubits: QES-TPE circuits reached 97-98% test accuracy with 60-75 params, matching NNs with $>$200 params.

Best-practice guidelines identified in (Nguyen et al., 2021) include:

• Fix single-qubit encoding and trainable rotations to $R_y$ gates,
• Use shallow ansatz depth (1–2) to avoid excessive noise and barren plateaus,
• Set entanglement $k \geq n$, incrementally increasing until validation accuracy plateaus,
• Use an initial random sample of $20$ embeddings, then $300$ surrogate-guided proposals, with $50$–$100$ full-trainings as a compute budget.
• For high-feature/low-qubit settings, reduce input dimensionality with a classical autoencoder prior to embedding.

4. Quantum Embedding Search: Surrogate Modeling and Optimization

The QES algorithm formulates embedding search as

$$\alpha^* = \arg\min_{\alpha\in\Omega_\mathrm{reduced}(k)} L_\mathrm{val}(\alpha),$$

where $L_\mathrm{val}$ quantifies performance (validation loss) of the QNN instantiated with entanglement pattern $\alpha$. The surrogate $S(\alpha)$ is trained on observed $(\alpha_i, \ell_i)$ pairs and is used to propose new candidates maximizing expected improvement (EI):

$$\mathrm{EI}(\alpha \mid M_{t-1}) = \int_{-\infty}^{\bar{t}} (\bar{t} - y)\,p_{M}(y \mid \alpha)\,dy,$$

where $p_M$ is the modeled distribution over losses. The surrogate-based approach drastically reduces the number of expensive full trainings required per run, leading to substantial computational efficiency, particularly critical in NISQ simulation regimes (Nguyen et al., 2021).

5. Expressivity, Parameter Efficiency, and Quantum Advantage

The QES framework enables discovery of quantum embedding circuits that are both parameter- and depth-efficient:

• For moderate $n$, QES-optimized circuits reach accuracy competitive with classical NNs or boosting methods with far fewer parameters.
• t-SNE visualization of quantum-encoded representations shows improved class separation, indicating the embedding circuit's ability to remap data geometry beneficially.
• Empirically, QES-TPE closes the performance gap to classical SVM/XGBoost, often with $\sim1/3$ the number of tunable weights and shallow circuit depths.

A key observation is that the quantum kernel $K(x,x')$ defined by the optimized embedding matches or surpasses classical kernels in class-separation capacity, confirming the utility of trainable quantum feature maps and the role of circuit structure search in practical QML (Nguyen et al., 2021).

6. Reproducibility, Implementation Trade-offs, and Outlook

QES and related quantum embedding workflows lend themselves to reproducibility on NISQ hardware or simulators:

• Each search consists of $T$ trials, each entailing (a) embedding proposal, (b) full circuit training, (c) validation, and (d) surrogate model update.
• Single run compute cost is governed by number of full trainings, with 4-qubit, depth-2 circuits on simulators consuming $2$–$4$ GPU-days per search at $k=4$ or $k=8$ entanglement.
• Shallow-layer, low-parameter embeddings are critical for robustness under hardware noise, and for maintaining tractable classical simulation times.

A plausible implication is that SMBO-driven quantum embedding search will remain a key component of NISQ-era hybrid QML practice, as hardware constraints and model selection demands outpace naive circuit enumeration or heuristic design.

7. Relation to Quantum Kernel Theory and Machine Learning Paradigms

Quantum embeddings form the theoretical bridge between quantum machine learning feature maps, quantum kernel methods, and entanglement structure optimization:

• The map $x\mapsto|\psi(x)\rangle$ is directly analogous to the nonlinear feature map in kernel SVMs,
• The QES framework generalizes classical "circuit architecture search" to the quantum domain, with direct impact on expressivity and generalization,
• The quantum kernel induced by the embedding provides access to analytically optimal measurements (e.g., Helstrom for trace-distance optimality),
• Tuning the mapping and associated circuit architecture controls metric learning capacity—maximizing between-class trace distance or Hilbert-Schmidt distance leads to empirically optimal minimum-risk classification (Nguyen et al., 2021, Lloyd et al., 2020).

The development of algorithmic quantum embedding search substantiates the critical insight that representational power in quantum ML arises from precise control over the mapping of input data to quantum state space, with efficient optimization algorithms serving as practical enablers in both simulation and hardware environments.