In classical deep learning, overparameterization is often a luxury we can afford. Adding extra layers and parameters might increase training time, but classical hardware executes these operations with near-perfect digital fidelity.
In the era of Noisy Intermediate-Scale Quantum (NISQ) computing, however, overparameterization carries a severe physical penalty. Every parameter in a Parameterized Quantum Circuit (PQC) or Variational Quantum Algorithm (VQA) corresponds to a physical quantum gate operation. Every gate we add exposes the system to environmental noise, state decoherence, and control inaccuracies.
To bypass this hardware limitation, a major research direction focuses on identifying and eliminating parameter redundancy. By systematically pruning unnecessary gates, researchers can compress quantum circuits, drastically reducing noise accumulation while preserving—or even improving—the circuit's expressivity.
The Geometry of Quantum Redundancy
Why do variational quantum circuits end up with redundant parameters in the first place? When designing an ansatz (the structural template of a quantum circuit), we rarely know the exact shape of the target quantum state. As a result, we rely on highly flexible, overparameterized templates to ensure the circuit is expressive enough to find a solution.
However, many of these parameters are geometrically redundant. In Dimensional Expressivity Analysis of Parametric Quantum Circuits, Lena Funcke et al. demonstrate that a parameter is redundant if the state changes it induces can be entirely replicated by adjusting other parameters in the circuit. In other words, these extra parameters do not expand the manifold of reachable quantum states; they merely create duplicate paths to the same states.
To find these redundancies on real, noisy hardware, Tobias Hartung et al. developed a method that analyzes the eigenvalues of state derivative overlaps (closely related to the Quantum Geometric Tensor). The intuition is elegant: if you wiggle a parameter and the resulting state change is linearly dependent on the changes from other parameters, it will manifest as a near-zero eigenvalue. The authors showed that even in the presence of hardware noise, truly independent parameters yield distinctively larger eigenvalues, allowing redundant gates to be cleanly isolated and removed.
Overparameterized Ansatz ──(Identify Redundant Dimensions)──> Minimal, Expressive Circuit
[Gate 1 (θ₁)] Analyze eigenvalues of [Gate 1 (θ₁)]
[Gate 2 (θ₂)] ──Redundant──> state derivative overlaps ──> [ PRUNED ]
[Gate 3 (θ₃)] to find duplicate paths [Gate 3 (θ₃)]
While Enrico Fontana et al. noted that overparameterization can occasionally help classical optimizers navigate around hardware errors during training, the consensus across the field is that physical circuit reduction is vital. As argued by the authors of Review on Quantum Computing for Lattice Field Theory, expressivity-based circuit optimization streamlines the optimization landscape and directly lowers the total noise floor.
How Quantum Pruning Differs from Classical Pruning
When we prune a classical neural network, we typically force unimportant weights to zero. In quantum circuits, the physics of rotational gates introduces a unique twist: circular geometry and periodicity.
A standard quantum rotation gate (like $R_x(\theta)$) acts as an identity operation—meaning it does absolutely nothing and can be deleted from the circuit—not just when $\theta = 0$, but also at periodic intervals like $2\pi$ or $4\pi$.
In Quantum Neural Network Compression, Zhirui Hu et al. introduced periodic pruning to exploit this phenomenon. Standard zero-centered pruning techniques fail to simplify gates that naturally settle near $2\pi$. By allowing parameters to prune at multiple periodic identity levels during training, their framework achieves significantly deeper circuit compression without sacrificing expressivity.
Classical Pruning: θ ───(Must drive to 0)───> Pruned
Quantum Pruning: θ ───(Can drive to 0, 2π, 4π...) ───> Reduces to Identity ───> Pruned
To prevent optimizers from wandering aimlessly through these infinite periodic valleys, Michael Kölle et al. proposed a technique called weight re-mapping. By mathematically constraining trainable parameters to a tight $[-\pi, \pi]$ interval, they eliminate the functional redundancy of identical angles, forcing the optimizer to take the most direct route to a solution.
To actively drive non-essential gates toward these identity points during optimization, researchers use sparsity-inducing mathematical priors:
- Laplace Priors: Samuel Duffield et al. applied Laplace priors within a Bayesian learning framework to automatically push non-essential gate angles to zero, improving the signal-to-noise ratio on physical hardware.
- L1 Regularization: In a framework called Rotolasso, Rudy Raymond et al. used $L_1$ regularization combined with coordinate descent to systematically turn off unnecessary gates, minimizing environmental decoherence.
Mitigating Training Overhead
Pruning a circuit after it has been fully trained solves the hardware execution problem, but it doesn't solve the training problem. Evaluating large, overparameterized quantum circuits requires millions of expensive physical measurements. If we have to fully train a giant circuit just to figure out what to prune, we have already wasted our computational budget.
To address this, Yifeng Peng et al. introduced TITAN, a framework that leverages the initial geometry of the quantum parameter landscape. They discovered that the circuit's starting state contains enough structural information to predict which parameters will remain inactive throughout the optimization process. By pre-emptively freezing these redundant dimensions before training even begins, they reduce measurement overhead by 40–60%.
Other approaches integrate pruning directly into the structural search of the circuit:
- Iterative Pruning: Hanrui Wang et al. developed QuantumNAS, which decouples circuit architecture search from parameter training. It iteratively removes gates with small parameter magnitudes and retrains the remaining structure to maximize noise resilience.
- Layerwise and Structural Reduction: Rather than evaluating individual gates, Jonas Stein et al. explored layerwise reduction patterns that prune gates symmetrically from the outermost qubits inwards. Similarly, Natacha Kuete Meli et al. embedded dedicated pruning layers into automated quantum architecture searches to discard low-impact operations dynamically.
By shifting from generic, bloated circuit templates to minimal, hardware-optimized structures, this collective body of work ensures that limited NISQ resources are spent entirely on meaningful quantum operations rather than redundant mathematical overhead.
Key Foundational Work
Adaptive pruning-based optimization of parameterized quantum circuits (2021) — contributed 3 concept elements to this direction.
Towards quantum simulations in particle physics and beyond on noisy intermediate-scale quantum devices (2021) — contributed 3 concept elements to this direction.
Quantum Neural Network Compression (2022) — contributed 3 concept elements to this direction.