The expressive capacity for learning with quantum systems is fundamentally limited by the quantum sampling noise incurred during measurement. While studies suggest that noise limits the resolvable capacity of quantum systems, its precise impact on learning remains an open question. We develop a framework for quantifying the expressive capacity of qubit-based systems from finite numbers of projective measurements, and calculate a tight bound on the expressive capacity and the corresponding accuracy limit that we compare to experiments on superconducting quantum processors. We uncover the native function set a finitely-sampled quantum system can approximate, called eigentasks. We then demonstrate how low-noise eigentasks improve performance for tasks such as classification in a way that is robust to noise and overfitting. We also present experimental and numerical analyses suggesting that entanglement enhances learning capacity by reducing noise in eigentasks. Our results are broadly relevant to quantum machine learning and sensing applications