Linear-time kernel goodness-of-fit test

• When reproducing the results, it will be necessary to modify the dictionary keys in kgof.config to have correct directories for various paths. If the keys are modified externally, then the same modifications need to be applied before plots can be generated by the provided Jupyter notebooks.

• When adding a new Kernel or new UnnormalizedDensity, use np.dot(X, Y) instead of X.dot(Y). autograd cannot differentiate the latter. Also, do not use x += .... Use x = x + .. instead.

• Note that current version of the test only works for cases where the domain is R^d (not even a bounded subset). This is due to the vanishing-boundary condition required to construct the Stein operator.

• The sub-module kgof.intertst depends on the linear-time two-sample test of Jitkrittum et al., 2016 (NIPS 2016). The code (freqopttest Python package) can be found here.