Multiple Choice

Why can a learning-curve point fluctuate when it is based on a very small random training subset?