Universal statistical laws governing culinary design

Problem
This preprint addresses the gap in understanding whether culinary recipes exhibit universal statistical laws akin to those observed in other symbolic systems, such as language. Despite the rich diversity of global cuisines, the authors investigate the underlying structural regularities in recipes, which have not been systematically quantified or compared to established statistical frameworks.

Method
The authors employ a large corpus of traditional recipes, which they annotate using a state-of-the-art named entity recognition (NER) algorithm. This annotation categorizes components into ingredients, cooking techniques, utensils, and other culinary attributes. The analysis reveals several statistical properties: ingredient usage follows a Zipf-like rank-frequency distribution, culinary diversity adheres to Heaps’ law (growing sublinearly with corpus size), and recipe complexity aligns with Menzerath-Altmann-type relations, indicating a relationship between the number of components and their average information content. Additionally, macronutrient concentrations across recipes exhibit a log-normal distribution. To model these findings, the authors propose minimal generative models that incorporate preferential reuse, constrained sampling, and incremental modification, which successfully replicate the observed statistical regularities.

Results
The study quantitatively demonstrates that ingredient usage adheres to Zipf’s law, with a clear rank-frequency relationship. Culinary diversity, as measured by the number of unique recipes, grows sublinearly, consistent with Heaps’ law, indicating that as the corpus expands, the rate of new unique recipes diminishes. The complexity of recipes, characterized by the relationship between the number of ingredients and their average information, follows Menzerath-Altmann-type relations, suggesting that more complex recipes tend to have a lower average information per ingredient. The log-normal distribution of macronutrient concentrations mirrors findings in packaged foods, reinforcing the statistical parallels. The generative models proposed effectively capture these dynamics, suggesting that simple, constrained processes can lead to complex culinary structures.

Limitations
The authors acknowledge that their analysis is limited to traditional recipes and may not generalize to modern or fusion cuisines, which could exhibit different statistical properties. Additionally, the reliance on NER for annotation may introduce biases or inaccuracies in categorizing culinary components. The study does not explore the cultural or contextual factors that may influence recipe construction, which could provide deeper insights into the observed statistical laws.

Why it matters
This work has significant implications for both culinary science and computational creativity. By establishing recipes as a compositional symbolic system governed by universal statistical laws, it opens avenues for further research into the generative processes underlying culinary design. The findings could inform the development of AI systems for recipe generation, enhancing their ability to create novel dishes that adhere to culturally relevant structures. Furthermore, this research contributes to the broader understanding of how complex systems can emerge from simple rules, a principle applicable across various domains in machine learning and artificial intelligence.

Authors: Ganesh Bagler, Gopal Krishna Tewari, Aditya Raj Yadav, Akshat Singh, Pranay Bansal, Ujjval Dargar, Mansi Goel, Madhvi Kumari Sinha
Source: arXiv:2604.28021
URL: https://arxiv.org/abs/2604.28021v1