Derivative Gaussian Processes on a Two-Direction Budget
The paper claims gradient data can be folded into Gaussian process prediction using only two directional derivatives per observed gradient.
The method compresses the usual \(md\) gradient coordinates into at most \(2m\) directions inside a Vecchia approximation. That keeps dense factorization at \(\mathcal{O}(m^3)\) per prediction target while still using gradient observations. The authors give error bounds against full-gradient conditioning and identify cases where the approximation is exact. In simulations, it matched a leading exact gradient-reduction method at equal conditioning size, then scaled to larger conditioning sets with less time and memory. ArXiv · AI/CL/LG's note
The method compresses the usual \(md\) gradient coordinates into at most \(2m\) directions inside a Vecchia approximation. That keeps dense factorization at \(\mathcal{O}(m^3)\) per prediction target while still using gradient observations. The authors give error bounds against full-gradient conditioning and identify cases where the approximation is exact. In simulations, it matched a leading exact gradient-reduction method at equal conditioning size, then scaled to larger conditioning sets with less time and memory. ArXiv · AI/CL/LG's note
score 4