Barren plateaus kill trainability in parameterized quantum circuits: gradients decay exponentially with system size, making optimization impossible. The standard explanation is observable concentration — the measurement operator averages to zero over the circuit's output distribution. Fix the observable, fix the gradient. Or so the field assumed.
Li et al. show that avoiding observable concentration is necessary but not sufficient. Two additional mechanisms suppress gradients independently.
Mid-circuit information loss: a parameter perturbation propagates through the circuit into degrees of freedom that the final measurement cannot access. The information exists — it just ends up in the wrong subsystem. The gradient vanishes not because the signal is diluted, but because it is redirected.
Mid-circuit information scrambling: a local perturbation spreads rapidly across the system and becomes effectively undetectable on the measured subsystem. The information is not lost to inaccessible degrees of freedom — it is smeared uniformly across accessible ones, becoming invisible against the background.
The three mechanisms are distinct. Observable concentration is about what you measure. Information loss is about where the signal goes. Information scrambling is about how the signal disperses. A circuit can avoid all three, or fail at any one independently. Quantum convolutional neural networks, for instance, exhibit information-loss-induced barren plateaus even with no observable concentration at all.
The unifying framework reveals that the barren plateau problem is deeper than its original diagnosis suggested. The gradient depends not just on the measurement and the circuit depth, but on the circuit's internal information flow — which subsystems talk to which, and whether perturbations stay localized long enough to be detected.