ConceptAdvancedTrainability

Barren Plateaus

Barren plateaus are the exponential vanishing, in the number of qubits, of the cost-gradient variance for deep random parameterized circuits, which makes training infeasible.

McClean, Boixo, Smelyanskiy, Babbush and Neven (2018) showed that if a parameterised circuit is random enough to approximate a 2-design on $n$ qubits, then the mean of the cost derivative is zero and its variance decreases exponentially with $n$: $$\mathbb{E}[\partial_kC]=0,\qquad \mathrm{Var}[\partial_kC]\in O(b^{-n}),\ b>1.$$ By Chebyshev's inequality, the probability that the derivative exceeds a small fixed value is exponentially small, so the landscape looks flat over most of parameter space.

The problem is not only the optimiser: to estimate a direction accurately enough we need a number of shots that grows exponentially, and this also affects gradient-free methods. Later work, including the review by Cerezo and colleagues (2021), showed that the phenomenon also depends on the cost function (local costs are less susceptible than global ones), on excessive entanglement, and on noise.

Proposed mitigations: shallow or problem-inspired circuits, local costs, smart initialisation, and layer-wise training. The effect does not show clearly in a simulation with few qubits.

Mathematical formula

$$\mathrm{Var}_\theta\left[\partial_k C(\theta)\right]\in O\!\left(b^{-n}\right),\quad b>1$$

Worked example

If the variance decreases roughly as $2^{-n}$, the standard deviation of the derivative is $2^{-n/2}$. For $n=4$: $0.25$; for $n=20$: $\approx0.001$; for $n=40$: $\approx10^{-6}$. To estimate a derivative of size $10^{-6}$ with shot error $1/\sqrt N$ we need about $10^{12}$ shots. (Base $2$ is used here for illustration; the actual value of $b$ depends on the circuit.)

Test your understanding

The quiz needs JavaScript.

Related topics

References

  1. Jarrod R. McClean; Sergio Boixo; Vadim N. Smelyanskiy; Ryan Babbush; Hartmut Neven (2018). Barren plateaus in quantum neural network training landscapes. Nature Communications 9:4812 doi:10.1038/s41467-018-07090-4
  2. M. Cerezo et al. (2021). Variational quantum algorithms. Nature Reviews Physics 3:625–644 doi:10.1038/s42254-021-00348-9