The following TR may answer some questions about how and why SVM's work so well: Optimal Properties and Adaptive Tuning of Standard and Nonstandard Support Vector Machines, by Grace Wahba, Yi Lin, Yoonkyung Lee and Hao Zhang, University of Wisconsin Statistics Dept TR 1045, October, 2001, available at ftp://ftp.stat.wisc.edu/pub/wahba/tr1045.ps Under some general circumstances flexible (kernelized) tuned appropriately, are implementing the Bayes Rule.