A concept mentioned by Nassim Nicholas Taleb in Fooled by Correlation: Common Misinterpretations in Social Science where he says "correlation is severely subadditive" and therefore "cannot be used for random subsamples"
He most likely borrowed his understanding of "subadditivity" from Tail Value at Risk (TVaR) in Finance, where:
1. TVaR is the expected value of losses when losses are greater than the Value at Risk i.e. E [ X | X >= VaR_α ] where α is the confidence level
2. TVaR -- unlike VaR -- satisfies "subadditivity", meaning combining portfolios never increases TVaR
3. TVaR -- unlike VaR -- accurately captures black swan events i.e. extremely rare events which add large quantities to the value losses