Dressed Quantum Network (DQN)

• DQN is a hybrid classical–quantum variational classifier that uses super compressed encoding to map high-dimensional data to a minimal qubit register.
• It employs only single-qubit gates for state preparation and variational transformations, significantly reducing noise and decoherence on NISQ devices.
• Benchmark results on datasets like Iris and WBC validate DQN’s competitive accuracy and scalability for efficient quantum machine learning.

The Dressed Quantum Network (DQN) is a hybrid classical–quantum variational classifier designed to address the limitations of variational Quantum Machine Learning (QML) on Noisy Intermediate-Scale Quantum (NISQ) hardware, specifically by reducing qubit requirements and minimizing noise sources. The DQN framework comprises a classical "super compressed encoding" layer followed by a compact quantum circuit composed exclusively of single-qubit gates and projective measurements. This architecture enables resource-efficient embedding of high-dimensional data into a minimal qubit register and leverages robust state preparation and variational transformations for effective supervised classification (Kumar et al., 2020).

1. Architecture and Data Flow

The DQN consists of two core components:

• Classical encoding layer: Implements a fully connected transformation $W\in\mathbb{R}^{d\times N}$, mapping the input vector $\mathbf{x}\in\mathbb{R}^d$ to an $N$-dimensional vector $\widetilde{\mathbf{x}}\in\mathbb{R}^N$, where $N$ is the number of target classes. The encoding is given by

$$\widetilde x_j = \sum_{i=1}^d w_i^j x_i\,,\quad j=1,\ldots,N.$$

• Quantum variational circuit: For each of the $N$ qubits,
  1. State preparation is performed by applying a $R_z(\widetilde x_j)$ rotation followed by a Hadamard gate:

  $$|\psi(\mathbf{x})\rangle = \bigotimes_{j=1}^N e^{i\sigma_3\widetilde x_j} H |0\rangle^j$$

1. Variational SU(2) transformation with parameters $$\boldsymbol\alpha^j = (\alpha_1^j, \alpha_2^j, \alpha_3^j)$$:

   $\mathbf{x}\in\mathbb{R}^d$0

2. Projective measurement onto the $\mathbf{x}\in\mathbb{R}^d$1 eigenstate yields the classwise probabilities $\mathbf{x}\in\mathbb{R}^d$2.

This structure decouples input dimension $\mathbf{x}\in\mathbb{R}^d$3 from qubit number $\mathbf{x}\in\mathbb{R}^d$4 and circumvents the need for multi-qubit gates, a significant source of decoherence and operational infidelity in NISQ devices.

2. Super Compressed Encoding

A distinguishing feature of DQN is the "super compressed encoding," which maps high-dimensional input $\mathbf{x}\in\mathbb{R}^d$5 to an $\mathbf{x}\in\mathbb{R}^d$6-tuple $\mathbf{x}\in\mathbb{R}^d$7, with $\mathbf{x}\in\mathbb{R}^d$8 in typical applications. This transformation both reduces the dimensionality and encodes data for direct quantum state preparation:

• Each component $\mathbf{x}\in\mathbb{R}^d$9 is utilized as the rotation angle in the $N$0 operation for qubit $N$1.
• The resulting compression allows implementations where the number of qubits required equals only the number of output classes, independently of the sample's original feature dimension.
• This aggressive reduction mitigates the hardware constraints of current quantum processors and enables scalable QML solutions on NISQ devices (Kumar et al., 2020).

3. Quantum Circuit Construction and Parameterization

The quantum portion of DQN comprises only single-qubit gates for each of the $N$2 qubits, structured as follows:

• State preparation: Each qubit is initialized by an $N$3 followed by a Hadamard.
• Variational layer: SU(2) gate sequence per qubit, decomposed as $N$4.
• Circuit depth: Per qubit, the sequence involves a single axial rotation, a Hadamard, and three more rotations.
• Parameters: The total quantum parameter count is $N$5, with all parameters being independent and trainable during circuit optimization.
• Measurement: Each qubit is measured in the computational basis, with class probabilities inferred from the resulting statistics (Kumar et al., 2020).

No two-qubit operations are present, which drastically reduces susceptibility to correlated errors and crosstalk.

4. Training Protocol and Optimization

DQN is trained using a standard supervised learning paradigm:

• Loss function: Multiclass tasks utilize cross-entropy loss:

$N$6

For binary classification, a linear loss function is applied:

$N$7

• Optimization: Classical optimizers such as Adam or RMS-prop are employed. Gradients with respect to all weights ($N$8) and quantum angles ($N$9) are computed via finite differences or automatic differentiation using PennyLane.
• No explicit regularization is used, but the restriction to single-qubit operations acts as a natural regularizer against circuit noise.
• The optimization loop iterates over the training set, accumulating loss, computing parameter gradients, and updating both classical and quantum parameters until convergence (Kumar et al., 2020).

5. Experimental Results and Hardware Implementation

The DQN was benchmarked on multiple datasets using diverse platforms:

Dataset	Classes	Python Sim.	Qiskit Sim.	IBM-Q
Fisher's Iris (d=4)	3	90%	94%	82%
WBC (d=30)	2	92.4%	96.5%	91.7%
Abalone (d=8)	6	67.7%	67.4%	67.2%

• Hardware platforms: Python (noise-free), Qiskit with PennyLane, and real IBM-Q devices (Rome, Armonk, Melbourne).
• Only single-qubit gates are needed; no CNOTs are required.
• Typical number of shots per circuit: 8,000–20,000.
• Reported single-qubit gate errors are between approximately 0.05% and 0.4% on IBM-Q, with measurement error around 1%.
• No active error-mitigation was used; however, simulations incorporating noise models were validated by comparison to experiment (Kumar et al., 2020).

6. Bloch-Sphere Clustering and Interpretability

For binary classification (single qubit case), DQN's operation can be interpreted via clustering on the Bloch sphere:

• Post-training, the classical encoding clusters the compressed inputs $\widetilde{\mathbf{x}}\in\mathbb{R}^N$0 such that the corresponding states $\widetilde{\mathbf{x}}\in\mathbb{R}^N$1 are mapped to two distinct arcs at $\widetilde{\mathbf{x}}\in\mathbb{R}^N$2 on the Bloch sphere.
• The variational SU(2) parameters then rotate these arcs toward the north ($\widetilde{\mathbf{x}}\in\mathbb{R}^N$3) or south ($\widetilde{\mathbf{x}}\in\mathbb{R}^N$4) pole, corresponding to output classes.
• The measurement probability for outcome $\widetilde{\mathbf{x}}\in\mathbb{R}^N$5 is given by $\widetilde{\mathbf{x}}\in\mathbb{R}^N$6; the role of the quantum variational layer is thus both separable and explainable via geometric visualization.
• The Bloch-sphere clustering provides direct intuition for the interplay between classical and quantum parameters and the mechanism of class separation (Kumar et al., 2020).

7. Significance and NISQ Suitability

DQN offers a variational QML solution tailored to NISQ hardware limitations:

• Requires only as many qubits as output classes, regardless of input dimensionality.
• Uses single-qubit gates exclusively, mitigating two-qubit gate-induced noise.
• Demonstrates competitive classification accuracy on standard benchmarks.
• Supports interpretability through geometric visualization on the Bloch sphere.

A plausible implication is that DQN's architectural simplicity and noise resilience may inform future designs for resource-constrained QML algorithms and foster further integration with NISQ-era hardware (Kumar et al., 2020).