In the current era of Noisy Intermediate-Scale Quantum (NISQ) computing, researchers face a "Goldilocks" dilemma: a quantum circuit must be expressive enough to reach a solution, yet shallow enough to avoid being overwhelmed by hardware noise. Most Variational Quantum Algorithms (VQAs) rely on human-designed templates, such as hardware-efficient or chemistry-inspired ansätze. However, recent research reveals that these templates are often "too fat"—they contain significant parameter redundancy that complicates optimization and invites decoherence.
Understanding and managing this redundancy has become a central pillar of quantum machine learning, shifting the focus from simply adding more gates to finding the minimal, most effective "winning tickets" within a circuit.
The Redundancy Paradox
At first glance, more parameters seem beneficial. In classical neural networks, overparameterization often smooths the optimization landscape. In quantum circuits, however, redundancy is a double-edged sword.
On one hand, Enrico Fontana, Ivan Rungger, and colleagues have shown in Evaluating the noise resilience of variational quantum algorithms that overparameterization can provide "noise-resilient" paths. By having more gates than strictly necessary, a classical optimizer can navigate around hardware errors to find parameter settings that maintain high fidelity.
On the other hand, this excess complexity is the primary driver of the "barren plateau" problem—a phenomenon where the gradient of the cost function vanishes exponentially, leaving the optimizer lost in a flat landscape. Xia Liu and Xin Wang argue in Mitigating barren plateaus of variational quantum eigensolvers that universal circuit architectures are inherently inefficient because they possess far more degrees of freedom than are required to represent a single target state. By restricting these degrees of freedom, they propose a State Efficient Ansatz (SEA) that reduces parameter counts quadratically, keeping the landscape trainable.
The Geometry of Quantum Landscapes
Why is there so much redundancy? The answer lies in the geometry of the parameter space. Enrico Fontana and Marco Cerezo explain in Optimizing parametrized quantum circuits via noise-induced breaking of symmetries that Parametrized Quantum Circuits (PQCs) possess an exponential number of hidden symmetries. These symmetries create a landscape with many identical global minima. While ideal simulators see these as equal, real hardware noise "breaks" these symmetries, turning some minima into "good" solutions and others into "bad" ones.
To quantify this, Tobias Hartung, Karl Jansen, and their team developed a framework called Dimensional Expressivity. As detailed in Determining the ability for universal quantum computing: Testing controllability via dimensional expressivity, they treat redundancy as a geometric property: a parameter is redundant if the change it induces in the quantum state can be perfectly replicated by adjusting other parameters. This allows researchers to map the "manifold of reachable operations" and identify which controls are actually pulling the same lever.
Pruning the Quantum Forest
If circuits are naturally redundant, can we simply cut away the parts we don't need? Several researchers are applying techniques from classical deep learning to do exactly that.
- The Lottery Ticket Hypothesis: Michael Kölle and Claudia Linnhoff-Popien demonstrated in Investigating the Lottery Ticket Hypothesis for Variational Quantum Circuits that VQCs contain sparse "winning ticket" sub-circuits. They found that pruning up to 74% of a circuit's parameters can result in a sparse model that performs as well as, or better than, the original dense circuit.
- Adaptive Freezing: Instead of pruning permanently, the TITAN framework developed by Yifeng Peng and Zhiding Liang predicts which parameters will be inactive during training. By "freezing" these redundant parameters early, they reduce measurement overhead by up to 60% in TITAN: A Trajectory-Informed Technique for Adaptive Parameter Freezing in Large-Scale VQE.
- Sparsity-Inducing Optimization: Rudy Raymond and Zichang He introduced Rotolasso, an optimization framework that uses coordinate descent to "turn off" unnecessary gates during the training process itself, as described in Optimizing Parameters of Quantum Circuits with Sparsity-Inducing Coordinate Descent.
Toward Gate-Frugal Architectures
The ultimate goal is to move away from generic, redundant templates toward "gate-frugal" designs. Mateusz Ostaszewski and Vedran Dunjko have pioneered the use of Reinforcement Learning (RL) agents to autonomously design these architectures. In Reinforcement learning for optimization of variational quantum circuit architectures, their agent discovered circuits for chemistry simulations that were significantly shallower than human-designed standards.
Some researchers suggest the problem lies in the gate abstraction itself. Zhiding Liang and Hanrui Wang argue in Towards Advantages of Parameterized Quantum Pulses that standard gates often have "parameter overlap," where different variables don't explore independent directions in Hilbert space. By shifting to pulse-level control, they bypass the overhead of gate abstractions entirely, achieving up to a 96% reduction in execution time.
By treating redundancy not just as a nuisance but as a geometric property to be mapped and pruned, the field is moving toward a more disciplined, resource-efficient approach to quantum algorithm design.