We derive sharp spectral-norm bounds for low-rank inverse approximations of noisy matrices, improving classical estimates by up to \sqrt{n} and offering spectrum-aware robustness guarantees validated on real and synthetic data.