Quantum coherence is usually a resource. Maintaining it is expensive — entire fields of quantum error correction and decoherence suppression exist because coherence is fragile and valuable. Noise destroys coherence, and destroying coherence should make quantum systems perform worse at quantum tasks.
Srivastava, Bhattacharyya, and Sen (arXiv 2602.16426, February 2026) find a quantum task where noise makes things better.
The task is nonlocal prediction. Alice and Bob share an entangled state. Alice performs a measurement on her subsystem. Without communicating the result, Bob tries to guess which outcome Alice got, using only his subsystem and knowledge of the shared state and measurement settings. For maximally entangled states, Bob can always guess perfectly — Alice's measurement creates orthogonal states on his side regardless of which basis she uses. For partially entangled states, the post-measurement states at Bob's end generally overlap, making perfect discrimination impossible.
Apply dephasing noise to one subsystem — randomly destroying the phase relationship between basis states — and for a broad class of partially entangled states and measurement settings, Bob's prediction accuracy increases. The dephased state outperforms the pure state.
The mechanism is precise. When Alice measures a partially entangled state in a basis rotated away from the natural Schmidt basis, the unequal Schmidt coefficients create quantum interference between the different entanglement components. This interference makes the post-measurement states at Bob's end overlap more than they would without coherence. The coherence is constructive for some purposes but destructive for state distinguishability — it aligns the post-measurement states, making them harder to tell apart.
Dephasing eliminates these off-diagonal coherences, converting the quantum superposition into a classical mixture. The classical mixture has no interference between components, and without the interference, Alice's measurement sorts the mixture into states that are more distinguishable at Bob's end. The noise doesn't add information. It removes coherence that was actively making the discrimination task harder.
The effect has specific boundaries. It vanishes for maximally entangled states, where Bob already achieves perfect prediction and there's nothing to improve. It vanishes when Alice measures in the Schmidt basis, where the pure state already produces orthogonal post-measurement states. It requires partial entanglement — unequal Schmidt coefficients that create the amplitude imbalance responsible for the detrimental interference — and measurement bases rotated away from the natural basis of the entanglement.
For two qubits, the condition reduces to an explicit inequality involving the entanglement parameter and the measurement rotation angle. When the rotation is small enough relative to the asymmetry in the Schmidt coefficients, dephasing helps. As the rotation increases, the advantage disappears. The boundary traces a curve in the parameter space that separates coherence-as-resource from coherence-as-liability.
The result doesn't contradict the general principle that coherence is valuable for quantum information processing. It identifies a specific regime where coherence creates harmful interference patterns, and noise, by eliminating those patterns, produces a net gain. The static doesn't carry signal. But sometimes the signal was distorting the picture.