El Libro de Arena
We formalise The Book of Sand (El Libro de Arena), a search game in which a human has a finite number of turns to find a hidden point in the countably infinite memory of an AI that answers every query with a hint. The score rewards both finding the point and the information gained about it, measured in bits. We prove that: without hints, or against an adversary, no finite budget suffices (the value of the game is zero); with a known prior, the optimal success probability is the mass of the n most likely points, rising to the 2^n − 1 most likely points with comparison hints; a randomly walking human finds the point almost surely in dimension d ≤ 2 but not in d ≥ 3 (Pólya's theorem); with noisy hints, the cost is governed by the channel capacity. The central result is an informational paradox. There are infinite-entropy memories in which the human finds the point almost surely and gains an unbounded amount of information, yet the fraction learned is zero at every finite turn and the expected search time is infinite. Four game levels with verified numerical values are given. This document is a mathematical support for a PhD thesis in applied robotics and artificial intelligence. It models the problem of an agent that must locate a target in an unbounded hypothesis space with a finite budget of actions: exploring an environment, retrieving an item from a very large memory, or querying a model to find out what it knows. Language: Spanish, with an English abstract.
Authors
- Javier Gutiérrez-Maturana Sánchez
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24428700
- Primary Topic
- Optimization and Search Problems
- Type
- preprint