Asymptotic mixture advantage for learning the optimum

Establish whether the asymptotic equivalence between mixture synthesis and individual synthesis, proved for learning the entire molecule-activity function, also holds when the objective is to learn the optimal molecule x⋆.

Background

The paper proves that, under a fully expressive molecule-activity model with Gaussian assay noise and information-gain-maximizing experimental design, testing mixtures and testing individual molecules yield the same asymptotic posterior covariance when the goal is to learn the entire activity function f. Thus, any advantage from information-dense stochastic synthesis is characterized as primarily a finite-data phenomenon for full-function learning.

The method LIDS, however, targets the maximizer x⋆ of the activity function rather than the entire function f. The authors explicitly note that their asymptotic result does not address this target and conjecture that analogous behavior may hold because x⋆ is a deterministic function of f. Proving or disproving this conjectured equivalence would clarify whether mixture synthesis can retain an asymptotic advantage specifically for optimum identification.

References

A limitation of this result is it applies to the EIG for learning the entire function $f_\theta$, while LIDS focuses just on the optimum $x\star$; we conjecture similar behavior for learning the optimum, since it is a deterministic function of $f_\theta$.

— Information-Dense Synthesis for Molecular Discovery  (2610.08495 - Jakobsen et al., 6 Oct 2026) in Section 5, Theory, subsection 5.2, immediately after Proposition 2 (the proposition labeled “Information bound” in the provided manuscript)