Optimal uncertainty bounds for multivariate kernel regression under bounded noise: A Gaussian process-based dual function

A researcher has derived a deterministic uncertainty bound for multi-output functions in Reproducing Kernel Hilbert Spaces that avoids the conservative estimates and restrictive noise assumptions of prior methods, according to a paper posted to arXiv in March 2026 and revised in May 2026 [1][2]. The work, authored by Amon Lahr, targets a persistent obstacle in safe learning-based control: existing uncertainty bounds for kernel methods such as Gaussian process regression are either overly conservative, require strong assumptions about noise distribution, fail in the multi-output setting, or resist integration into downstream optimization tasks [1][2]. The new bound is obtained through an unconstrained, duality-based formulation that mirrors the structure of classic Gaussian process confidence bounds, allowing it to be inserted directly into optimization pipelines [2]. Kernel methods have long been valued for their built-in uncertainty quantification. Gaussian process regression, for instance, models latent functions by assuming that any finite collection of function values follows a multivariate normal distribution, a property that yields closed-form posterior means and variances [3]. The normal distribution's analytic tractability — linear combinations of independent normal deviates remain normal, and least-squares parameter fitting can be derived explicitly — has made it a default modeling choice in many scientific domains [3]. However, real-world noise rarely conforms exactly to Gaussian assumptions, and when the noise is merely bounded rather than normally distributed, classical confidence intervals can misrepresent the true uncertainty [2]. The paper generalizes existing results to the multi-output case, a setting where multiple correlated functions must be estimated simultaneously. Multi-output problems arise naturally in applications such as quadrotor dynamics learning, which the author uses as an illustrative example [2]. In such systems, state variables like position, velocity, and orientation evolve jointly, and a single-output bound applied independently to each channel would ignore cross-channel dependencies, potentially underestimating risk. The proposed bound addresses this by operating directly in the RKHS of vector-valued functions [2]. Dimensionality reduction techniques such as principal component analysis are sometimes paired with kernel methods to manage high-dimensional multi-output data by projecting onto directions of maximum variance [4]. While the paper does not employ PCA, the broader numerical analysis toolkit — spanning optimization, linear algebra, and approximation theory — provides the mathematical substrate on which RKHS-based bounds are constructed [5]. The revised manuscript, dated 24 May 2026, runs to 773 KB, up from the initial submission's 757 KB [1]. The bound's tightness and deterministic nature mean it does not rely on probabilistic tail bounds that inflate uncertainty estimates, a feature the author argues makes it suitable for safety-critical control where over-conservatism can render a controller inoperable [2].