Barren plateaus are just a training trick to fix

Barren plateaus are fundamentally different from classical vanishing gradients. The gradient variance vanishes exponentially in the number of qubits — not just as a function of depth. For a global cost function on n qubits, the gradient magnitude scales as O(2^-n). No classical optimization trick fixes this because it is a consequence of the geometry of high-dimensional Hilbert space. Local cost functions partially mitigate this but introduce other trade-offs between trainability and expressiveness.

This concept requires either mathematical proof or hardware-scale experiments beyond what a browser simulator can demonstrate. See the research notes for the canonical references.