pith. sign in

arxiv: 2605.05967 · v1 · submitted 2026-05-07 · 💻 cs.LG · math.OC· stat.ML

Sharper Guarantees for Misspecified Kernelized Bandit Optimization

Pith reviewed 2026-05-08 14:21 UTC · model grok-4.3

classification 💻 cs.LG math.OCstat.ML
keywords misspecificationgammasqrtamplificationboundsofflineonlinesetting
0
0 comments X

The pith

For kernels with monotone spectra or Fourier-diagonal structure, misspecification penalties in kernelized bandits reduce to log or polylog factors via spectral and spatial localization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Kernelized bandits use a mathematical similarity function to guess the best action from past data. When the true reward function does not exactly match the kernel model, previous theory said the mismatch error would grow with the square root of how complex the kernel is. This work shows that for many practical kernels the growth can be slowed to logarithmic by focusing on local parts of the space or on the kernel's frequency properties instead of global complexity. The offline version uses a spectral Lebesgue constant; the online version splits the domain so local errors do not spread.

Core claim

For a large class of kernels, the misspecification amplification can be reduced to logarithmic or polylogarithmic growth. In the offline setting, high-probability simple-regret bounds whose misspecification term is governed by a spectral Lebesgue constant; in the online setting, cumulative regret of O~(sqrt(gamma_n n) + n epsilon) under mild localized eigendecay assumptions.

Load-bearing premise

The kernels belong to the class with one-dimensional monotone spectra or multivariate Fourier-diagonal product kernels, and the online analysis requires mild localized eigendecay assumptions that allow domain splitting to control global amplification.

read the original abstract

Existing guarantees for misspecified kernelized bandit optimization pay for misspecification through kernel complexity: in generic offline bounds, the misspecification level $\varepsilon$ is multiplied by $\sqrt{d_\mathrm{eff}}$, where $d_\mathrm{eff}$ is the kernel effective dimension, while in online regret bounds, the corresponding penalty is $\sqrt{\gamma_n}\,n\varepsilon$, where $\gamma_n$ is the maximum information gain after $n$ rounds of interaction. In this work, we show that, for a large class of kernels, the misspecification amplification can be reduced to logarithmic or polylogarithmic growth. In the offline setting, we first prove high-probability simple-regret bounds whose misspecification term is governed by a spectral Lebesgue constant. This yields logarithmic amplification for one-dimensional monotone spectra and polylogarithmic amplification for multivariate Fourier-diagonal product kernels. In the online setting, we modify a domain-splitting algorithm and prove a cumulative regret bound of $\widetilde{\mathcal O}(\sqrt{\gamma_n n}+n\varepsilon)$ under mild localized eigendecay assumptions, removing the extra $\sqrt{\gamma_n}$ factor from the misspecification term. The common principle is localization: spectral localization controls the Lebesgue constant of the offline approximation operator, while domain splitting implements the spatial analogue of this mechanism in the online setting, preventing local misspecification errors from being amplified globally.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only view limits visibility into exact assumptions; the central claims rest on standard kernel RKHS properties plus the paper-specific localized eigendecay and monotone/Fourier-diagonal spectrum conditions.

axioms (2)
  • domain assumption Kernels admit spectral decomposition with monotone or Fourier-diagonal structure
    Invoked to control the Lebesgue constant and enable localization.
  • domain assumption Mild localized eigendecay holds
    Required for the online domain-splitting regret bound.

pith-pipeline@v0.9.0 · 5560 in / 1254 out tokens · 34617 ms · 2026-05-08T14:21:44.968478+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.