Neural-Network Error Mitigation for Quantum Teleportation and Superdense Coding
Hritik Chaudhary, K. C. Gokul
Abstract
Abstract Quantum teleportation and superdense coding are foundational protocols of quantum communication, but both degrade under channel imperfections. This paper examines whether a small neural network, acting purely as a classical post-processing layer, can mitigate such imperfections when it is given only the information a real receiver could collect. Both protocols are implemented exactly using tensor operations on complex qubit amplitudes, reproducing ideal noiseless behaviour. Realistic imperfections are then introduced and learned mitigation is evaluated under finite-shot conditions, so that the network observes a finite number of discrete measurement outcomes rather than exact probabilities, and is never supplied the noise parameters. The results identify a clear boundary on when learning helps. For superdense coding through a biased Pauli channel followed by asymmetric readout error, a learned decoder exceeds a majority-vote baseline in the low-shot regime, by roughly five percentage points at five shots, with both methods converging toward perfect accuracy as shots grow. For teleportation through a systematic gate miscalibration, the network learns to invert the miscalibration and raises mean state fidelity from about 0.93 to 0.99, with the gain increasing as additional shots reduce tomographic noise. A control experiment replacing the systematic miscalibration with a fully random per-qubit rotation yields no benefit, the expected outcome when noise carries no recurring structure. The findings indicate that lightweight learned post-processing is worthwhile specifically in the low-shot, structured-noise regime, and offers nothing when noise is structureless.