The post on my blog, has much more details on the math of how to bridge the attention operator (a nonlinear operator)'s approximation capacity with a linear operator (Petrov-Galerkin projection). The ...