Barren plateaus and trainability
Training a variational circuit means walking downhill on an energy landscape. In 2018 Jarrod McClean and colleagues at Google showed that for generic, expressive circuits the landscape is almost perfectly flat: gradients are exponentially small in the number of qubits, everywhere except in a vanishing fraction of parameter space. To see the slope, you need exponentially many measurements. They called these regions barren plateaus, and they reshaped the field.
On the plate: the spread of gradients in random circuits, as qubits are added.
Why the landscape is flat
A circuit deep and random enough behaves like a random unitary (technically, a 2-design). Then the cost concentrates around its average, just as the overlap of two random vectors in a huge space concentrates near zero. The mean gradient is zero and its variance shrinks like or faster. This is concentration of measure, the same effect that makes random high-dimensional vectors nearly orthogonal. Established
Global versus local cost
The plate shows gradient variances computed by exact simulation of a layered circuit of random , rotations and controlled-Z gates, with as many layers as qubits. Show , , . A global cost asks whether the whole register is in ; its gradient variance falls from 0.037 at 2 qubits to 1.2 × 10⁻⁶ at 10. A local cost averages single-qubit measurements; it falls from 0.026 to 2.5 × 10⁻⁴, far more slowly. Marco Cerezo and colleagues proved in 2021 that local costs with shallow circuits avoid barren plateaus. Samson Wang and colleagues showed the same year that hardware noise causes plateaus of its own, whatever the cost function. Established
The twist: trainable may mean simulable
In 2024 two groups, Michael Ragone and colleagues and Enrico Fontana and colleagues, tied gradient variance to the dimension of the Lie algebra generated by the circuit’s gates: a small algebra means no plateau. But a small algebra also means the dynamics lives in a small subspace. In 2025 Cerezo and colleagues argued that most provably plateau-free models can then be simulated classically, possibly after collecting some data from the quantum device. If true in general, the structure that makes a model trainable is the same structure that removes its quantum advantage. Exceptions, and how far the argument extends, are actively debated. Debated
What you now have
Barren plateaus as concentration of measure, the exponential decay of gradient variance, the advantage of local costs, noise-induced plateaus, the Lie-algebraic theory, and the trainability–simulability tension.
· · ·
Three questions to carry with you.
And one to work out. Suppose gradient variance halves with each added qubit, and the shots needed scale as 1 / variance. By what factor do shots grow from 10 to 20 qubits?
Sources: J. R. McClean et al., Nat. Commun. 9, 4812 (2018). M. Cerezo et al., Nat. Commun. 12, 1791 (2021). S. Wang et al., Nat. Commun. 12, 6961 (2021). M. Ragone et al., Nat. Commun. 15, 7172 (2024). E. Fontana et al., Nat. Commun. 15, 7171 (2024). M. Cerezo et al., Nat. Commun. 16, 7907 (2025). Checked by code: exact state-vector simulation, 120–300 random circuits per size, parameter-shift gradients of the first angle.