betaFeedback

WorksPhysics and Astronomy

Higher-order statistical structure in perfume item combinations beyond a pairwise maximum-entropy model

Abstract

We find that deviations from a pairwise maximum-entropy model in perfume item combinations are organized around particular shared item pairs. We analyze 1,979 perfumes using 50 binary variables representing the most frequently listed items. Although the pairwise model approximately reproduces the distribution of the number of listed items per perfume, the observed three-item marginal distributions exhibit deviations exceeding those obtained from synthetic datasets subjected to the same model-refitting procedure. Strongly deviating triplets are concentrated around shared item pairs, and this concentration cannot be explained by the participation frequencies of individual items alone. We then introduce three-body couplings for 50 selected triplets and reoptimize all model parameters. The sum of squared standardized residuals over the remaining 19,550 triplets decreases by 5.44%, with substantially larger reductions for triplets sharing an item pair with selected triplets. Furthermore, a static linear-response calculation accounting for the readjustment of lower-order parameters approximately reproduces the resulting changes in predicted co-occurrence frequencies, with a Pearson correlation coefficient of 0.850. These results reveal the organization of higher-order statistical dependencies and characterize how selected three-body couplings affect predictions throughout the system.

Paper details

vlnc:W_cecktshu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
  • Physics and Astronomy

ChatGPT (OpenAI), Claude (Anthropic)

View statement
Looking for related papers…