Barren plateaus: why variational quantum algorithms are hard to train

Short answer: A barren plateau is a flat, featureless region of a variational quantum algorithm's training landscape where gradients vanish, which makes the circuit hard to train as it grows. A 2025 Nature Reviews Physics review says choices of ansatz, initial state, observable, loss function and hardware noise can all cause it. A 2021 paper proves that, for local Pauli noise, gradients vanish exponentially if circuit depth grows linearly with qubit count. This is an open research problem, not a solved one.

What causes the optimization landscape to flatten?

The primary mechanism behind barren plateaus is the exponential flattening of the parameter optimization landscape as the number of qubits or layers increases. When a model exhibits this behavior, the landscape becomes featureless, meaning there are no significant gradients to guide the optimizer toward a better solution. This issue is not limited to a single component of the algorithm. Instead, all moving pieces—including the choice of ansatz, initial state, observable, loss function, and hardware noise—can contribute to the formation of barren plateaus if they are ill-suited. Consequently, the trainability of the algorithm is strongly impacted by how these components interact. The 2025 review calls barren plateaus a key obstacle to realizing the potential of variational quantum computing.

How does noise specifically induce barren plateaus?

Noise on Noisy Intermediate-Scale Quantum (NISQ) devices can place fundamental limitations on variational quantum algorithm performance. For local Pauli noise, rigorous proofs show that the gradient vanishes exponentially in the number of qubits if the depth of the ansatz grows linearly with the number of qubits. These noise-induced barren plateaus are conceptually different from noise-free barren plateaus, which are typically linked to random parameter initialization. The theoretical result is formulated for a generic ansatz, including special cases such as the Quantum Alternating Operator Ansatz and the Unitary Coupled Cluster Ansatz. For one ansatz, the authors say their numerical heuristics demonstrate the effect for a realistic hardware noise model.

What strategies exist to mitigate the problem?

To address the hindrance that barren plateaus pose to scaling variational quantum circuits on large datasets, extensive efforts have been devoted to developing mitigation strategies. A 2025 survey proposes a taxonomy that categorizes existing mitigation strategies into five distinct groups. These strategies aim to tackle the issue from both investigation and mitigation perspectives. The study of barren plateaus has become a thriving area of research, influencing and exchanging ideas with other fields such as quantum optimal control, tensor networks, and learning theory. By understanding the theoretical underpinnings, researchers can develop heuristic methods to mitigate the effects of these plateaus, although the scalability to large problem sizes remains the main open question for variational quantum algorithms.

Why is this considered a research-stage challenge?

Variational quantum algorithms are currently viewed as a hybrid computational approach aiming to train quantum learning models. While they offer a flexible approach with a broad range of applications, the barren plateau phenomenon remains a key obstacle to realizing their full potential. Methods to understand and mitigate the effects are being developed, and the problem is not resolved. We make no claim of successful large-scale training or production use.

FAQ

What is a barren plateau?

A phenomenon in which a model's parameter optimization landscape becomes exponentially flat and featureless as the problem size increases, so there is little gradient to guide training.

What causes it?

A 2025 review says all the moving pieces of an algorithm can lead to barren plateaus if ill-suited: the ansatz, initial state, observable, loss function and hardware noise.

How does noise cause barren plateaus?

A 2021 Nature Communications paper proves that for the local Pauli noise it considers, the gradient vanishes exponentially in the number of qubits if the ansatz depth grows linearly with it. It calls these noise-induced barren plateaus.

Are there fixes?

A 2025 survey proposes a taxonomy grouping existing mitigation strategies into five groups. The review describes the study of barren plateaus as a thriving research area.

Does this mean variational algorithms work at scale?

The sources describe barren plateaus as a key obstacle. This article makes no claim of large-scale or deployed use.