Trainability and Barren Plateaus in QML

Module 7 introduced barren plateaus for variational algorithms in general. Here we focus on what they mean specifically for QML models — classifiers, kernels, and generative circuits — where the plateau interacts with the data encoding in ways that pure ground-state problems do not show.

The core result, restated for QML

For a sufficiently expressive (roughly, $2$ -design-like) parameterized circuit on $n$ qubits, the gradient of the cost with respect to any parameter concentrates around zero, with variance shrinking exponentially in the qubit count:

\operatorname{Var}\!\left[\frac{\partial \mathcal{L}}{\partial \theta_k}\right] \in \mathcal{O}\!\left(\frac{1}{2^{\,b\,n}}\right), \qquad b > 0.

When this holds, the landscape is an exponentially flat plateau: a gradient estimated from a realistic number of shots is indistinguishable from noise, so the optimizer wanders aimlessly. The phenomenon was identified by McClean et al. (2018).

QML-specific sources of plateaus

In a learning model the cost is not arbitrary — it is built from data and a readout. That structure opens several distinct plateau mechanisms:

Expressibility-induced. Too-expressive an ansatz makes the circuit look like a random unitary, triggering the original McClean mechanism. More expressive is not always better.
Cost-function-induced. Global observables (acting on all qubits at once) give plateaus even at shallow depth, whereas local cost functions — readouts on a few qubits — can stay trainable. This is the central message of Cerezo et al. (2021): make the loss local.
Entanglement-induced. Excessive entanglement between the "visible" readout qubits and the rest drives the reduced state toward maximally mixed, flattening gradients.
Encoding-induced. Heavy data encoding (many features, many re-uploads) can itself randomize the state and concentrate gradients — a uniquely QML route to a plateau.

Kernels concentrate too

Quantum kernel methods are not immune. Thanasilp et al. (2022) showed that the kernel entries

k(\boldsymbol{x},\boldsymbol{x}') = \bigl|\langle \phi(\boldsymbol{x})|\phi(\boldsymbol{x}')\rangle\bigr|^2

can exponentially concentrate: for expressive feature maps on many qubits, almost all off-diagonal entries collapse toward the same tiny value, so every pair of distinct points looks equally dissimilar. The kernel matrix becomes nearly diagonal, the data is no longer separable in feature space, and exponentially many shots are needed just to resolve the entries. The plateau problem of variational models thus has an exact analogue in the kernel picture.

Mitigations carried into QML

The defenses from Module 7 specialize cleanly to learning models:

Local cost functions / local readouts instead of global observables.
Problem-informed, shallow ansätze — including symmetry-preserving and hardware-efficient-yet-structured designs.
Warm starts and layerwise training so early layers begin in a trainable region.
Restrained encoding — fewer re-uploads and lower-dimensional feature maps when concentration threatens.

The takeaway

Trainability is the gatekeeper of QML. A model can be expressive, elegant, and still learn nothing because its gradients or its kernel entries have vanished into the noise floor. Before trusting any QML result, check that the gradients are resolvable at the chosen qubit count and shot budget — the benchmarking lessons that follow build exactly this discipline into the evaluation.