Dynamic Parameterized Quantum Circuits

• DPQCs are quantum circuits that dynamically adjust structure and parameters via in-circuit measurements and feed-forward control, enabling enhanced expressibility and adaptation.
• They incorporate intermediate measurements and classical logic to select subcircuits adaptively, improving performance in generative modeling, optimization, and quantum machine learning.
• DPQCs mitigate barren plateaus through dynamic parameter tuning and efficient gradient scaling, offering robust optimization for complex quantum algorithms.

Dynamic Parameterized Quantum Circuits (DPQCs) are quantum circuit architectures that combine parameterized unitaries with in-circuit measurements, classical feed-forward, and—often—adaptively conditioned subcircuit selection. This class generalizes standard parameterized quantum circuits (PQCs) by enabling circuit action, topology, or parameter values to change dynamically based on measurement outcomes and additional control information, dramatically expanding expressibility, learnability, and practical utility across generative modeling, optimization, and quantum machine learning.

1. Definition and Structural Features

DPQCs are architectures in which, in contrast to static PQCs (which have fixed circuit structure with tunable rotation angles), the circuit action at each stage may depend on the outcomes of previous measurements or on ancillary registers. Formally, the output quantum channel implemented by a DPQC is generally described as an ensemble:

$$\mathcal{E}(\rho) = \sum_{i} p_i \cdot U_i(\theta_i) \cdot \rho \cdot U_i(\theta_i)^\dagger$$

where branch index $i$ is selected dynamically via measurement and/or post-selection, and $U_i$ is a unitary (or controlled unitary) acting with parameters $\theta_i$ that may themselves depend on measurement outcomes, prior state preparation, or classical data.

Principal mechanisms include:

• Intermediate measurement and classical control: The ability to perform measurements mid-circuit and condition further operations (e.g., apply a gate $G$ if and only if a previous measurement returns bit value $b$).
• Ancilla-driven control and post-selection: The joint system-ancilla state is prepared, manipulated via conditional gates, and either post-selected or used to induce effective subcircuit selection (e.g., for Bayesian learning).
• Dynamically adapted subcircuits: Subcircuit blocks, with either different parameterizations or structure, are selected based on previous measurement registers or classical control flow.
• Dynamic parameter tuning: Dynamic reparameterization of gates at runtime based on real-time input (e.g., feedback from previous outcomes or across distributed quantum processors).

These features support both measurement-outcome-conditioned channels and classical logic integration, resulting in a dynamic computational graph for circuit execution (Du et al., 2018, Hong et al., 2021, Vazquez et al., 27 Feb 2024).

2. Expressive Power and Entanglement Structure

A haLLMark of DPQCs is their superior expressive power for generative modeling when compared to both classical neural network architectures and static (non-dynamic) PQCs:

• Volume-law entanglement: DPQCs (and especially multilayer PQCs with ancillary control or post-selection) can efficiently generate quantum states and associated probability distributions with volume-law scaling of the entanglement entropy $S(\rho_A) = O(|A|)$ for a subsystem $A$. In contrast, classical models such as short-range RBMs are restricted to area-law scaling $S(\rho_A) = O(|\partial A|)$ (Du et al., 2018).
• Complexity-theoretic separation: MPQCs can efficiently simulate the output statistics of IQP circuits:

$$p_I = |\langle 0^{\otimes N} | H^{\otimes N} U_Z H^{\otimes N} | 0^{\otimes N} \rangle|^2$$

where $U_Z$ is a diagonal unitary constructed from polynomially many commuting gates. No classical neural network can efficiently sample from this distribution unless the polynomial hierarchy collapses. Ancilla-driven and post-selection-enabled DPQCs further extend capacity by enabling conditional sampling from distributions that are at least as hard as post-IQP outputs.

Summary Table: Expressive Power Ordering

Model variant	Entanglement scaling	Simulates IQP?
MPQC / DPQC	Volume law ($O(|A|)$)	Yes
Deep Boltzmann	Volume law	No (unless DBM with global connections)
Long-range RBM	Volume law	No
Tensorized PQC	Area law	No
Short-range RBM	Area law	No

The dynamic features (ancilla-driven control, measurement-based branches) therefore not only increase expressive capacity but establish DPQCs as strictly more powerful than deep classical neural networks, barring a collapse of the polynomial hierarchy (Du et al., 2018).

3. Optimization and Gradient Landscape

Optimization in DPQCs inherits the challenges and opportunities of PQCs—particularly the barren plateau phenomenon—while introducing distinctive mechanisms to address these issues:

• Fourier-based derivative estimation: For MiNKiF-type circuits (which generalize standard rotation-parameterized gates), one may estimate gradients by evaluating the circuit’s expectation value

$f(t) = F(\theta + t e_j)$

at several points $t_1, \ldots, t_S$ in parameter space and reconstructing the derivative via Fourier inversion without introducing ancillas or additional circuit complexity. For broader classes of Hamiltonians, this exploits the Fourier expansion

$$f(t) = \frac{1}{\sqrt{2\pi}} \sum_k \hat{f}(k) e^{ikt}$$

and recovers the derivative as

$$f'(0) = \frac{1}{\sqrt{2\pi}} \sum_k (ik) \hat{f}(k)$$

with the support $D = \{k_i - k_j\}$ defined by the Hamiltonian spectrum (Vidal et al., 2018).

• Optimization trade-offs: The use of multiple evaluation points increases per-step cost but provides a trigonometric polynomial representation of the objective function along each parameter direction, enabling nonlocal updates (such as full coordinate descent steps) and enhanced robustness for highly dynamic parameter landscapes as in DPQCs.
• Gradient scaling and effective parameterization: For rotation-generated DPQCs, direct calculation shows the gradient expectation is not strictly zero but decays as $1/3^n$ for $n$-qubit circuits (Yao et al., 7 Mar 2025). The gradient variance behaves as $m/(8^n n d)$ for $m$ effective parameters and depth $d$, indicating that increasing the fraction of effective parameters can mitigate barren plateau effects. This improves upon earlier “Haar-based” analyses and supports careful ansatz design and pruning—removal of redundant parameters based on QFI eigenvalues—to optimize trainability in deeply dynamic architectures (Haug et al., 2021, Yao et al., 7 Mar 2025).
• Barren-plateau free architectures: DPQCs that incorporate measurement and feed-forward can provably avoid barren plateaus, as the structure of measurement-induced randomness prevents the exponential suppression of gradient variance characteristic of static PQCs (Deshpande et al., 8 Nov 2024).

4. Algorithmic and Generative Applications

DPQCs are particularly effective in applications requiring high representational flexibility and adaptation to incomplete, time-varying, or hierarchical information:

• Bayesian and semi-supervised learning: Bayesian quantum circuits (BQC)—a DPQC architecture using ancilla registers to encode latent variables—can generate both priors and class-conditional distributions. By optimizing over both data and ancilla blocks, these circuits learn class priors and posteriors directly from data, extending quantum generative model performance and accuracy in settings (e.g., class-imbalanced data) where priors are unknown (Du et al., 2018). In numerical experiments, DPQC-based BQCs reach >99% accuracy in low-dimensional tasks and can learn nonuniform priors with low variance.
• Continuous generative modeling: Expectation-value sampling-based DPQCs can universally approximate continuous multivariate distributions. Two key circuit designs are distinguished:
  • Product encoding (one qubit per variable, $O(M)$ qubits), enabling efficient measurement due to bounded observable norms.
  • Dense encoding (logarithmic qubits, $O(\log M)$), which demands observables with size scaling as $O(M)$ and leads to a trade-off: fewer qubits at the cost of increased measurements for the same additive accuracy (Barthe et al., 15 Feb 2024).
• Physics-informed learning and Sobolev spaces: DPQCs, when designed as truncated Fourier series approximators with suitable data normalization, can approximate functions not only in $L^2$ but also in $H^k$ Sobolev spaces. Loss functions incorporating all derivatives up to order $k$ can thus be minimized, granting precise control for quantum models of differential equations and physical processes (Manzano et al., 2023).

5. Implementation: Hardware, Verification, and Real-Time Adaptation

• Hardware acceleration and runtime reconfiguration: Modern FPGA-based controllers can identify structurally equivalent circuits (that differ only by parameters), “peel off” variable gate angles for waveform reuse, and stitch them in real time at the hardware level (e.g., QubiC and its Stitch module). This permits execution of large batches of DPQCs with different parameterizations at compilation speedups up to $995\times$ and classical time reductions exceeding 85% (Rajagopala et al., 5 Sep 2024). Such techniques enable dynamic parameter variation and circuit adaptation crucial for DPQC workflows.
• Distributed, cross-QPU dynamic circuits: Real-time classical communication channels allow measurement results from one quantum processor to control operations on another, enabling scalable execution of dynamic circuits with modular qubit resources, circuit-cutting, and nonplanar connectivity. Staggered dynamical decoupling and zero-noise extrapolation are essential for robust execution of such dynamically parameterized operations (Vazquez et al., 27 Feb 2024).
• Formal verification: Equivalence checking for DPQCs can be posed via tensor decision diagrams (TDDs), capable of encoding both parameterized unitary operations and classically controlled logic in a canonical, space-efficient representation. Both measurement (m-equivalence) and quantum state (q-equivalence) criteria can be used to compare circuits, with adaptation possible for parameterized/symbolic circuits (Hong et al., 2021).

6. Complexity, Geometry, and Structural Optimization

• Complexity scaling and topology: DPQCs reach Haar-like complexity—measured by expressibility (Kullback–Leibler divergence), average entanglement (Mayer–Wallach measure), and Lorenz curve fluctuations (majorization)—with fewer gates than random universal circuits, especially in well-connected topologies (linear, ring). This efficient complexity build-up supports both rapid mixing and resource-efficient quantum computation (Correr et al., 29 May 2024).
• Quantum geometry and pruning: The quantum Fisher information (QFI) matrix quantifies circuit expressibility (effective quantum dimension $G_C$) and identifies redundant parameters. Pruning algorithms remove such redundancy without sacrificing expressiveness, reducing circuit depth and measurement overhead (Haug et al., 2021).
• Structure and parameter co-optimization: Jointly optimizing both the continuous rotation angles and the generator (axis) for each gate—using sinusoidal parameter dependence and analytic minimizers—permits rapid, low-variance energy minimization in variational quantum algorithms. The “Rotoselect” and “Rotosolve” procedures enable shallower circuits with equivalent or better expressivity, critical for noisy devices (Ostaszewski et al., 2019).

7. Challenges, Limitations, and Recent Developments

• Barren plateau mitigation: The exponential suppression of Fourier coefficients in the output function of PQCs persists in dynamic settings whenever the underlying blocks form unitary 2-designs. This indicates that certain DPQC architectures are still subject to vanishing gradients, regardless of classical parameter initialization (Okumura et al., 2023). However, architectures that systematically introduce measurements and classical control can escape this limitation and provably avoid barren plateaus (Deshpande et al., 8 Nov 2024).
• Optimization landscape and effective parameters: Analytical and numerical work confirms that the gradient expectation in rotation-layer DPQCs decays as $1/3^n$ with $n$ qubits but is not strictly zero; the effective parameter count must be considered with respect to circuit depth to preserve gradient variance in deep variational circuits (Yao et al., 7 Mar 2025).
• Generative design and automation: Diffusion models can generate DPQC architectures and continuous parameters simultaneously from textual or performance-based prompts. This generative approach enables efficient sampling over diverse circuit ansätze optimized for target fidelity or classical accuracy, opening new directions in automated quantum circuit design and architecture search (Barta et al., 27 May 2025).

---

Current research indicates that DPQCs, by leveraging mid-circuit measurement, adaptive control, and dynamically updated parameters, form a highly expressive, versatile, and scalable class of quantum circuits well-suited to the demands and constraints of near-term quantum hardware. Their practical implementation now spans robust classical-quantum optimization, scalable hardware execution, and advanced generative design frameworks, with theoretical results solidifying their role in universal function approximation, efficient complexity generation, and barren plateau avoidance.