TODO list

[SebaGraz]

• [x] Implement the exact subsampling algorithm where ∇ϕ(x) is estimated by an unbiased estimator ∇ϕ_tilde(x) (the standard example is Big Data with i.i.d observations. Then the target density is given by a product of terms -> gradient of the negative log density is given by a sum of terms. Subsampling by picking randomly one term of the sum.

[mschauer]

• [x] Scaling in the 1-d Boomerang
• [x] Local n-d Boomerang

[SebaGraz]

I want to propose the following:

1. To have the d dimensional implementation as the default implementation and drop the 1-dimensional implementation
2. To group the implementations for not factorised (Bouncy and Boomerang) and factorised pdmps (Zig-Zag and Fact. Boomerang).
3. To have the Local implementation (with the PriorityQueue) of the factorised pdmps as the default implementation for those algorithms (Zig-Zag and Fact. Boomerang). When the graph is not specified, then assume a fully connected graph

[SebaGraz]

• [x] Save a and b and use them for both the evaluation of \lambda_bar and poisson_time. (Now we compute them twice)

[mschauer]

Let's keep the 1-d version intact for didactical purposes, but the high dimensional should be the default

[mschauer]

• [x] Allow for change of reference measure in FactBoomerang