Using statistical physics, we study support vector machines (SVMs) learning noisy target rules in cases when the optimal predictor is a polynomial of the inputs. If the kernel of the SVM has sufficiently high order or is transcendental, the scale of the learning curve and the asymptote is...