Variational quantum algorithms optimize a parameterized circuit by computing gradients of a cost function and adjusting gate parameters accordingly. The barren plateau problem — where gradients vanish exponentially with system size — is a central obstacle: if gradients become exponentially small, no optimizer can find the direction to improve, and the algorithm fails. Noise-induced barren plateaus extend this problem to noisy hardware: even circuits that avoid barren plateaus in the noiseless case were predicted to develop exponentially vanishing gradients when run on real quantum processors, because noise accumulates with circuit depth and washes out the gradient signal.
Schmitt, Ekstrom, Bottarelli, and Bonet-Monroig (arXiv 2602.22851, February 2026) run experiments on IBM quantum hardware with up to 102 qubits and find that noise-induced barren plateaus do not occur under the dominant noise channel of their devices.
The key distinction is the type of noise. The theoretical prediction of noise-induced barren plateaus assumed depolarizing noise — a unital channel that drives the quantum state toward the maximally mixed state. Under depolarization, the output becomes progressively closer to the identity as noise accumulates, and the gradient decays exponentially toward zero. But the dominant noise on superconducting qubit hardware is amplitude damping — a non-unital channel characterized by the T1 coherence time, which drives the state toward the ground state rather than the maximally mixed state.
Under amplitude damping, the gradient does not decay exponentially. It saturates. Beyond a characteristic circuit runtime set by the T1 time, the gradient magnitude reaches a constant value that does not decrease further with increasing circuit depth or system size. The gradient is reduced from its noiseless value — the noise does suppress the signal — but the suppression is bounded, not exponential. The optimizer has a diminished but usable gradient at any circuit depth.
The saturation occurs because non-unital noise has a fixed point that is not the maximally mixed state. Depolarizing noise drives everything toward complete ignorance — the state where all information is erased. Amplitude damping drives everything toward the ground state — a specific, structured state that retains information about the circuit. The gradient with respect to early gates survives because the asymptotic state is not featureless. It has structure that the early gates influence.
The experimental demonstration uses Information Content Landscape Analysis to estimate gradient norms without directly computing gradients — a technique that extracts gradient information from the statistics of measured expectation values. The analysis spans circuits with 8 to 102 qubits, confirming the saturation behavior across two orders of magnitude in system size.
The result does not solve the barren plateau problem. It removes one predicted obstacle — noise-induced gradient vanishing — from the list, under the specific condition that the dominant noise is non-unital. Hardware with significant depolarizing noise (from crosstalk, leakage, or coherent errors) may still exhibit noise-induced barren plateaus. The distinction between noise types, usually treated as a technical detail of error modeling, turns out to determine whether variational quantum computing is fundamentally feasible on a given platform.