Boundary Variance Inflation Causes Acquisition Bias in Gaussian Processes
The paper shows that Gaussian-process variance can inflate near boundaries, biasing acquisitions away from the objective. It also proposes a function-free diagnostic to measure the effect.
Intelligence analysis by GPT-5.4 Mini
This arXiv paper argues that the geometry of bounded domains can distort Gaussian-process acquisition behavior on its own. The authors trace the issue to truncated correlation neighborhoods at the boundary and show that different acquisition rules are pushed toward corners or inward shells, even without any objective signal.
The paper says a smart guessing tool can get confused near the edge of a play area, like a flashlight that seems brighter at the walls because the room cuts off its beam. That can make it pick spots for the wrong reason, not because those spots are actually best.
Analysis
Core claim
The paper studies Gaussian processes with stationary kernels on bounded domains and argues that the familiar increase in posterior variance near a boundary is not just a nuisance detail. According to the abstract, the root cause is geometric: the kernel’s correlation neighborhood gets cut off at the edge of the domain, and that distortion becomes worse as dimensionality rises.
What this does to acquisition
The authors say this boundary effect can dominate several acquisition rules. For variance maximization, selections concentrate at corners. For negative integrated posterior variance and expected predictive information gain, selections move inward instead, forming axis-aligned interior shells. The important point is that these patterns appear without any objective function in the loop, so the acquisition rule can be steered by kernel geometry rather than by the task-specific uncertainty a user actually wants.
Main contribution
To study the phenomenon more generally, the paper introduces a function-free selection-profile diagnostic. The abstract says this diagnostic works for arbitrary acquisitions, kernels, and bounded-domain geometries. That makes the analysis portable: it is not tied to one optimization objective or one particular application setting.
Takeaway
The paper frames boundary variance inflation as a structural source of acquisition bias in Gaussian-process methods. For Bayesian optimization and related workflows, that means the choice of domain shape and kernel geometry can matter as much as the acquisition formula itself. The work is positioned as a way to understand, measure, and compare these effects more systematically.
Key points
- Gaussian processes on bounded domains can show inflated posterior variance near boundaries.
- The paper attributes the effect to truncated kernel neighborhoods at the edge of the domain.
- Different acquisition rules are biased in different ways: corners for variance maximization, inward shells for two other criteria.
- These patterns can appear without any objective function signal.
- The authors introduce a function-free diagnostic to study selection profiles across kernels and geometries.
If the paper’s diagnosis holds up, researchers can better spot when Gaussian-process acquisitions are being pulled around by boundary geometry instead of by useful uncertainty. The proposed diagnostic could help compare kernels, acquisition rules, and domain shapes more fairly.
If this effect is ignored, acquisition methods may keep over-sampling corners or interior shells for the wrong reason, especially in higher dimensions. That can waste evaluations and distort Bayesian optimization results even when the model itself looks mathematically sound.
