Parameterized Quantum Circuits (PQCs)

• Parameterized quantum circuits (PQCs) are quantum circuits with fixed and parameter-dependent gates that serve as variational ansätze for hybrid quantum algorithms.
• They are characterized by metrics like expressibility (measured via KL divergence against Haar randomness) and entangling capability (quantified by the Meyer–Wallach measure).
• Design principles emphasize layered architectures, noise mitigation, and hardware-aware optimizations to enhance performance in quantum machine learning, optimization, and chemistry applications.

A parameterized quantum circuit (PQC) is a quantum circuit composed of both fixed and parameter-dependent gates, forming a family of unitaries $U(\theta)$ acting on an $n$-qubit state. These circuits underpin the majority of variational quantum algorithms (VQAs)—notably for quantum machine learning, combinatorial optimization, and quantum chemistry—where their tunable parameters $\theta$ are optimized by a classical outer loop to extremize an objective defined as an expectation value or classifier loss. The design, expressibility, trainability, and robustness to noise of PQCs have been the subject of extensive research and rigorous quantitative analysis.

1. Mathematical Structure and Variational Role

A PQC on $n$ qubits is defined by

$$U(\theta) = U_L(\theta_L) \cdots U_2(\theta_2)U_1(\theta_1)$$

where each $U_j(\theta_j)$ is either a single-qubit rotation $\exp(-i\theta_j \sigma_\alpha/2)$ (for $\alpha \in \{x, y, z\}$), or a multi-qubit entangling gate (e.g., CNOT, CZ), with $\theta = (\theta_1,\dots,\theta_m)$ the set of trainable parameters. A typical PQC instance prepares a state $$|\psi(\theta)\rangle = U(\theta)|0^{\otimes n}\rangle$$ and measures an observable $M$ to yield the cost function $$C(\theta) = \langle\psi(\theta)|M|\psi(\theta)\rangle$$ (Benedetti et al., 2019, Aktar et al., 2023, Benítez-Buenache et al., 2024).

In hybrid quantum-classical algorithms (e.g., VQE, QAOA, quantum classifiers, and QCBM), the classical optimizer proposes updates to $\theta$ based on observed or estimated losses, closing the feedback loop. The PQC thus acts as the variational ansatz, whose expressive power determines the reachable hypothesis space and ultimately the performance on learning or optimization tasks (Benedetti et al., 2019).

Commonly, the circuit is organized in a layered fashion, with each layer consisting of a parallel block of parameterized single-qubit rotations and entangling gates, repeated $L$ times for expressibility control. Modular design allows arbitrary feature encoding (e.g., amplitude or angle encoding) to precede the parameterized block, as is standard in quantum machine learning (Jones et al., 10 Jul 2025, Benítez-Buenache et al., 2024).

2. Expressibility, Entangling Capability, and Descriptors

Expressibility measures the circuit's ability to generate a diverse state ensemble spanning Hilbert space, compared to the Haar measure. Formally, for $|\psi(\theta)\rangle$ drawn uniformly at random over the parameter domain, sample the squared overlap $F=|\langle\psi(\theta)|\psi(\theta')\rangle|^2$, gather its empirical distribution $P_\mathrm{PQC}(F)$, and compare to the analytic Haar distribution

$P_\mathrm{Haar}(F) = (2^n - 1)(1 - F)^{2^n - 2}.$

The scalar expressibility metric is the Kullback–Leibler divergence:

$$\mathrm{Expr} = D_\mathrm{KL}[P_\mathrm{PQC}(F)\,\|\,P_\mathrm{Haar}(F)] = \int_0^1 P_\mathrm{PQC}(F)\log\frac{P_\mathrm{PQC}(F)}{P_\mathrm{Haar}(F)}\,dF$$

with $\mathrm{Expr}\to0$ indicating near–Haar randomness (Sim et al., 2019, Aktar et al., 2023, Liu et al., 2024, Azad et al., 2022). Expressibility can also be characterized by frame potentials $\mathcal{F}^{(t)}$:

$$\mathcal{A}^{(t)}(C) = \mathcal{F}^{(t)}_C - \mathcal{F}^{(t)}_\mathrm{Haar}$$

where $C$ is the PQC and $\mathcal{F}^{(t)}$ is computed as the $t$-th moment of the state overlap (Liu et al., 2024).

Entangling capability is quantified by the Meyer–Wallach measure:

$$Q(|\psi\rangle) = \frac{2}{n} \sum_{j=1}^n (1 - \mathrm{Tr}\,\rho_j^2)$$

where $$\rho_j = \mathrm{Tr}_{\neq j}[|\psi\rangle\langle\psi|]$$. The empirical average over sampled $\theta$ yields a global entangling capability score (Sim et al., 2019, Azad et al., 2022).

Empirical studies demonstrate that expressibility and entangling capability "saturate" with increased circuit depth or two-qubit gate count, with additional layers yielding diminishing improvements (Sim et al., 2019, Liu et al., 2024). Circuit topology (e.g., all-to-all, block-ring, ring, line) and entangling gate type (CRX vs. CRZ) systematically affect both expressibility and entanglement (Liu et al., 2023, Sim et al., 2019, Liu et al., 2024).

3. Circuit Design Principles and Topological Optimization

The architecture of a PQC—choice of entanglers, gate axes, block structure, and two-qubit layout—directly determines its expressibility and trainability. Systematic studies have found:

• All-to-all entangling layouts maximize expressibility and entanglement but incur quadratic $O(n^2)$ two-qubit gate cost.
• Block-ring ("BR") topologies reduce gate count to $O(mn)$ (for tunable block size $m$) while achieving near–all-to-all performance in shallow circuits. This interpolation gives almost-maximal metrics after only two layers (Liu et al., 2023).
• Choosing CRX gates yields higher expressibility and entanglement than CRZ, due to non-commuting two-qubit interactions (Sim et al., 2019, Liu et al., 2023).
• Parameterized rotations about arbitrary Bloch-sphere axes (Fraxis), or free quaternions (FQS), provide a richer search space and improved expressibility compared to fixed-axis schemes such as $R_y$ or discrete axis selection (Watanabe et al., 2021, Pankkonen et al., 10 Jul 2025).
• Expressibility increase from adding RX or RY gates saturates for normalized counts $\gtrsim10$–15 per qubit; excess CNOTs may reduce expressibility due to entangling–rotation imbalance. Practical design rules recommend $$(\text{RX}+\text{RY})\!:\!\text{CNOT} \simeq 3\!:\!1$$ for maximal coverage (Liu et al., 2024).

Hardware constraints, including connectivity, error rates, and native gate sets, require customized PQC designs for different physical platforms, often via hardware-aware architecture search (e.g., Bayesian optimization over topology and gate type) (Benítez-Buenache et al., 2024). For combinatorial optimization, dynamically constructed PQCs that explicitly encode problem constraints into amplitude support greatly reduce search space and yield faster convergence compared to hardware-efficient, unconstrained ansätze (Matsuo et al., 2020).

4. Training, Optimization, and Information-Theoretic Aspects

PQC training is performed using measurement-based evaluation of the objective function, which may be the expectation of an observable (VQE), cross-entropy for classification (QML), or a statistical divergence for generative modeling. Updates of $\theta$ are made via:

• Gradient-based methods, using the parameter-shift rule:

$$\frac{\partial C}{\partial\theta_k} = \frac{C(\theta_k+\pi/2) - C(\theta_k-\pi/2)}{2}$$

provided $U(\theta_k)$ is generated by a reflection operator. These allow for classical optimizers (SGD, Adam, natural gradient), with update steps $\theta\leftarrow\theta-\alpha\nabla C(\theta)$ (Pankkonen et al., 9 Oct 2025, Benedetti et al., 2019, Li et al., 2024).

• Gatewise coordinate-descent optimizers:
  • Rotosolve: updates one angle at a time, fitting a trigonometric model for $C(\theta_d)$ (Pankkonen et al., 9 Oct 2025, Pankkonen et al., 10 Jul 2025).
  • Free-Axis Selection (Fraxis): optimizes both axis and angle for each single-qubit rotation, efficiently via a 3×3 eigenproblem when $\theta=\pi$, achieving KL expressibility $\lesssim 0.01$ (Watanabe et al., 2021).
  • Free Quaternion Selection (FQS): upgrades the search to arbitrary SU(2) elements, further enlarging the search space (Pankkonen et al., 9 Oct 2025).
• Hybrid optimization protocols: Rotosolve–FQS or Rotosolve–Fraxis hybrids, controlled by early stopping or cost-improvement thresholds, combine rapid initial descent with global expressibility for deeper convergence and improved noise robustness (Pankkonen et al., 9 Oct 2025).

Information-theoretic analyses reveal that exact expectation-value queries are highly informative, while single-shot samples carry exponentially little information, setting sample complexity lower bounds for faithful training: $\Omega(2^n)$ shots are generally required to resolve "barren plateaus" and identify optimal regions, unless problem structure, prior information, or advanced measurement strategies (e.g., parameter-shift rule, grouped measurements, classical shadows) are exploited (Dolzhkov et al., 2019).

Gradient-vanishing and barren-plateau phenomena, where the objective gradient decays exponentially with system size or circuit depth, reliably appear for highly expressive or random ansätze. Strategies that restrict the variational subspace, increase locality, or adaptively extend the ansatz in response to stagnation (e.g., nested quantum-classical optimization with quantum-gradient feedback) have been shown to mitigate these effects and maintain trainability on polynomial resources (Li et al., 2024, Matsuo et al., 2020).

5. Noise Robustness and Error Mitigation

On realistic Noisy Intermediate-Scale Quantum (NISQ) devices, PQC performance is limited by decoherence, cumulative gate errors, and readout noise, which are exacerbated by deep and highly entangled circuits (Nguyen et al., 2024, Jones et al., 10 Jul 2025).

• Noise Accumulation: Each PQC layer accumulates uncorrected physical noise, causing the output state to progressively resemble a maximally mixed state—a process directly analogous to classical diffusion toward complete noise (Nguyen et al., 2024).
• Diffusion-Inspired Mitigation: Mitigation algorithms interleave noise channels with learnable "denoiser" steps, training a forward–backward protocol that regularly inverts the effect of noise. This approach, by adding a round-trip divergence loss to the task loss, consistently delivers 1–2 percentage-point improvements in classification accuracy over previous state-of-the-art noise mitigation on standard quantum learning benchmarks (Nguyen et al., 2024).
• Architecture Search Under Noise: Bayesian circuit architecture search (BPQCO) in the noise-aware setting yields shallower, more robust PQCs that balance expressibility and complexity, with Pareto-frontiers between performance and gate-error-induced complexity. Multi-objective optimization of test accuracy and transpiled circuit complexity has been shown to yield the best noisy results (Benítez-Buenache et al., 2024).
• Resource Optimization: Gate-freezing methods, which track the magnitude of parameter updates and temporarily exclude converged gates from further optimization, reduce the number of circuit evaluations by up to an order of magnitude while maintaining or improving convergence and final performance, as empirically demonstrated on Heisenberg and Fermi–Hubbard models (Pankkonen et al., 10 Jul 2025).

6. PQCs as Machine Learning Models and Universality

PQCs serve as flexible, physically motivated models for quantum machine learning tasks, ranging from regression and classification of classical or quantum data to generative modeling and unsupervised learning (Benedetti et al., 2019, Du et al., 2018, Jones et al., 10 Jul 2025). Notable theoretical results and architecture benchmarks include:

• Approximation Universality: PQCs can approximate any square-integrable, continuous, or Sobolev-class function given sufficient circuit depth and suitable data encoding, with formal $L^2$, $L^p$, $C^0$ and $H^k$ universal approximation theorems. Generalization bounds link empirical loss on samples and derivatives to target norm distances in function space, under appropriate normalization and parameter regularization (Manzano et al., 2023).
• Expressive Power Beyond Classical Models: Multilayer PQCs (MPQCs) with $O(\mathrm{poly}(n))$ blocks exhibit greater expressive power than classical generative neural networks (e.g., RBMs, DBMs), unless the polynomial hierarchy collapses. Ancilla-driven and post-selection–enabled circuits (Bayesian Quantum Circuits, BQC) can efficiently simulate classically intractable distributions and implement Bayesian learning by encoding priors in ancillary degrees of freedom (Du et al., 2018).
• Application Benchmarks: PQCs have been experimentally deployed for variational quantum classifiers (Iris/MNIST), quantum circuit Born machines for generative modeling, quantum autoencoders, and molecular regression in quantum chemistry. Empirical results demonstrate feasibility of chemically meaningful regression on NISQ devices using shallow PQCs with error mitigation, though challenges remain in scaling depth and qubit number simultaneously due to data–depth–noise trade-offs (Jones et al., 10 Jul 2025, Benedetti et al., 2019).

Empirical benchmarks stress the importance of matching ansatz expressibility and structure to the specific data or Hamiltonian, balancing depth, entangling power, and noise tolerance (Jones et al., 10 Jul 2025).

7. Visualization, Diagnostics, and Design Tools

Advanced visualization and diagnostic tools for PQCs—such as qLEET—provide capabilities for:

• Quantifying expressibility and entangling power landscape over ansatz families.
• Visualizing high-dimensional loss landscapes and projective slices, revealing ridges, valleys, and barriers responsible for optimization bottlenecks.
• Mapping out training trajectories of the optimizer within parameter space, showing convergence dynamics and basin structure.
• Using the entanglement spectrum (eigenvalues of the reduced density matrix) to assess proximity to randomness (Marchenko–Pastur law) and to link ansatz structure to entropic properties (Azad et al., 2022).

Integrating such diagnostics with automated expressibility predictors (e.g., GNN regressors trained on circuit graphs (Aktar et al., 2023)) enables rapid pruning of under-expressive designs, guiding the choice of hyperparameters, gate patterns, and circuit depth prior to resource-intensive quantum training. Empirical evidence supports the utility of these methods for practical PQC architecture selection under tight resource and noise constraints (Aktar et al., 2023, Azad et al., 2022).

---

References

• (Benedetti et al., 2019) Parameterized quantum circuits as machine learning models
• (Du et al., 2018) The Expressive Power of Parameterized Quantum Circuits
• (Liu et al., 2023) A Block-Ring connected Topology of Parameterized Quantum Circuits
• (Sim et al., 2019) Expressibility and entangling capability of parameterized quantum circuits for hybrid quantum-classical algorithms
• (Liu et al., 2024) Analysis of Parameterized Quantum Circuits: on The Connection Between Expressibility and Types of Quantum Gates
• (Watanabe et al., 2021) Optimizing Parameterized Quantum Circuits with Free-Axis Selection
• (Pankkonen et al., 10 Jul 2025) Gate Freezing Method for Gradient-Free Variational Quantum Algorithms in Circuit Optimization
• (Pankkonen et al., 9 Oct 2025) Enhancing Hybrid Methods in Parameterized Quantum Circuit Optimization
• (Benítez-Buenache et al., 2024) Bayesian Parameterized Quantum Circuit Optimization (BPQCO): A task and hardware-dependent approach
• (Azad et al., 2022) qLEET: Visualizing Loss Landscapes, Expressibility, Entangling Power and Training Trajectories for Parameterized Quantum Circuits
• (Jones et al., 10 Jul 2025) Parametrized Quantum Circuit Learning for Quantum Chemical Applications
• (Aktar et al., 2023) Predicting Expressibility of Parameterized Quantum Circuits using Graph Neural Network
• (Matsuo et al., 2020) Enhancing VQE Convergence for Optimization Problems with Problem-specific Parameterized Quantum Circuits
• (Nguyen et al., 2024) Diffusion-Inspired Quantum Noise Mitigation in Parameterized Quantum Circuits
• (Dolzhkov et al., 2019) Information content of queries in training Parameterized Quantum Circuits
• (Manzano et al., 2023) Approximation and Generalization Capacities of Parametrized Quantum Circuits for Functions in Sobolev Spaces
• (Li et al., 2024) Learning Parameterized Quantum Circuits with Quantum Gradient