In a support vector machine (SVM), bias is systematic error caused by a boundary too simple to capture the underlying class structure. A small penalty parameter C can tolerate more slack and preserve a wider margin, increasing the risk of high bias and underfitting. A large C can make the boundary overly responsive to training observations, increasing the risk of high variance and overfitting. These are tendencies, not guarantees; no value of C is universally best. section 8.3.3
The objective shows how C weights slack against the margin term, setting the tradeoff that shapes the boundary. section 8.3.3





































