In the Noisy Intermediate-Scale Quantum (NISQ) era, quantum computing faces a fundamental engineering bottleneck: the trade-off between circuit expressivity and physical fidelity. To represent complex quantum states—such as molecular wavefunctions in chemistry or high-dimensional data distributions in machine learning—an algorithm requires high expressivity, which naturally demands a large number of sequential entangling gates (high circuit depth).
However, as highlighted by Enrico Fontana et al. in Evaluating the noise resilience of variational quantum algorithms, every additional gate layer exposes the system to environmental decoherence and gate-level errors. If a circuit's depth exceeds the hardware's coherence time, the delicate quantum state degrades into classical noise before a meaningful measurement can occur.
This tension is pervasive across near-term applications. Whether attempting to scale quantum layers for differential equations (Pengpeng Xiao et al. in Quantum DeepONet: Neural operators accelerated by quantum computing) or executing data-uploading schemes for machine learning (Maniraman Periyasamy et al. in Incremental Data-Uploading for Full-Quantum Classification), researchers consistently hit a wall where increasing depth destroys computational accuracy. To make near-term quantum computing viable, a concerted research direction has emerged to compress, optimize, and rethink quantum circuits to achieve maximum expressivity at minimal depth.
Offloading Complexity to Classical Infrastructure
The most prominent paradigm for bypassing depth constraints is the hybrid quantum-classical architecture. Instead of executing a massive, monolithic quantum algorithm, the computation is split: the quantum processor prepares a parameterized state and measures its properties, while a classical computer handles the heavy lifting of optimization.
This concept originated with the Variational Quantum Eigensolver (VQE), introduced by Alberto Peruzzo et al. in A variational eigenvalue solver on a photonic quantum processor. By shifting the optimization loop to classical hardware, VQE drastically reduces the continuous coherence time required from the quantum processor.
To push this further, researchers have developed methods to trade circuit depth for other classical resources. For instance, Hsin-Yuan Huang et al. proposed the Classical Combination of Variational Quantum States (CQS) framework in Near-term quantum algorithms for linear systems of equations with regression loss functions. Instead of building a single deep circuit to represent a complex state, CQS prepares multiple shallow quantum states and combines them classically, effectively trading circuit depth for an increased number of measurement repetitions. Similarly, Philipp Schleich introduced Quantum Deep Equilibrium Models to avoid the depth accumulation of multi-layered variational circuits by finding fixed points through classical root-finding algorithms rather than deeper physical layering.
Mathematical and Hardware-Native Reformulations
Another way to minimize depth is to change the mathematical "language" used to frame the problem, ensuring that the target state can be represented natively with fewer gates.
In quantum chemistry, standard variational templates often require gate counts that scale poorly with system size. To combat this, Panagiotis Kl. Barkoutsos et al. in Quantum algorithms for electronic structure calculations: Particle-hole Hamiltonian and optimized wave-function expansions shifted the mathematical starting point to a particle-hole representation, achieving high chemical accuracy with significantly shallower circuits. In a similar vein, Joonho Lee et al. developed the $k$-UpCCGSD ansatz in Generalized Unitary Coupled Cluster Wave functions for Quantum Computation, which uses a sparse approach to electron excitations to achieve a circuit depth that scales linearly with system size.
Beyond chemistry, optimization algorithms can be compressed by tailoring them to the underlying hardware physics. In Ion-native variational ansatz for quantum approximate optimization, the authors demonstrated that utilizing the native entangling operations of trapped-ion processors prevents the compilation overhead and depth explosion that occurs when forcing a hardware-agnostic algorithm onto a physical chip. Furthermore, the authors of Choco-Q: Commute Hamiltonian-based QAOA for Constrained Binary Optimization embedded mathematical constraints directly into the quantum driver via commute Hamiltonians, eliminating the exponential circuit depth typically required by standard penalty-based methods.
Automated Architecture Search
Because human intuition often struggles to find the absolute most efficient gate layout, researchers have turned to automated optimization frameworks to eliminate redundant operations and enforce strict depth constraints.
Quantum Architecture Search (QAS) leverages machine learning to discover "gate-frugal" circuit designs. As outlined by Samuel Yen-Chi Chen et al. in Introduction to Quantum Machine Learning and Quantum Architecture Search, QAS automates the discovery of shallow, noise-resilient configurations.
[Classical Optimizer / RL Agent]
│
▼ (Proposes Gate Sequence)
┌────────────────────────────────┐
│ Shallow Quantum Circuit │ ──► [High Fidelity Output]
└────────────────────────────────┘
▲ (Evaluates Noise/Depth)
│
[Hardware Constraints / Penalties]
For example, Mateusz Ostaszewski et al. trained reinforcement learning agents in Reinforcement learning for optimization of variational quantum circuit architectures to construct circuits that are significantly shallower than human-designed standards while maintaining chemical accuracy. To optimize both discrete gate selection and continuous parameters simultaneously, Haoya Li et al. introduced a hybrid framework combining Monte Carlo Tree Search with natural policy gradients in Monte Carlo Tree Search based Hybrid Optimization of Variational Quantum Circuits, yielding shorter, highly noise-resilient gate sequences.
Extracting Signal via Mitigation and Ensembles
When circuit depth cannot be reduced any further, researchers use error mitigation and ensemble methods to extract clean computational signals from shallow, noisy states.
Instead of relying on full quantum error correction, which requires massive qubit overhead, Shi‐Xin Zhang et al. introduced a "variational post-selection" framework in Variational post-selection for ground states and thermal states simulation. By pairing shallow circuits with ancilla qubits and classical neural network reweighting, they dramatically boost the expressive power of the circuit without increasing its physical depth.
Alternatively, noise can be managed at the algorithmic level using classical machine learning meta-algorithms like boosting. In Supervised Learning Guarantee for Quantum AdaBoost, Yabo Wang et al. proved that an ensemble of multiple "weak" (shallow and noisy) quantum classifiers can be combined classically to exponentially reduce training error. This ensemble approach allows a collection of shallow, hardware-compatible circuits to collectively outperform a single, ideally noiseless deep circuit that current hardware cannot reliably execute.