testing by betting—two-sample testing

For two-sample testing, Shekhar and Ramdas use $S_t=1+{\lambda }_t(X_t-Y_t)$ (more generally $S_t=1+{\lambda }_t(k(X_t)-k(Y_t))$ if $X_t,Y_t$ are multivariate observations, and $k$ is some kernel.

todo