The radial basis function (RBF) kernel lets a support vector machine compare observations by distance, enabling localized nonlinear decision boundaries. section 8.6.2
The kernel similarity decreases exponentially with the squared distance between observations; γ controls how quickly it falls. section 8.6.2
The fitted boundary depends on parameters as well as the kernel:
- A smaller γ spreads each support vector’s influence over a wider region, generally producing a smoother boundary; a larger γ makes that influence more local and can lead to overfitting. section 8.6.3
- γ interacts with the soft-margin penalty C. Evaluate candidate settings with cross-validation rather than selecting them from training accuracy alone. section 8.6.3
Comparing kernel boundaries
On the same synthetic two-class data, the RBF kernel produces a localized nonlinear boundary, unlike the linear kernel’s straight boundary. section 8.6.2



