Abstract
Mathematical proofs are both paradigms of certainty and some of the mostexplicitly-justified arguments that we have in the cultural record. Their veryexplicitness, however, leads to a paradox, because their probability of errorgrows exponentially as the argument expands. Here we show that under acognitively-plausible belief formation mechanism that combines deductive andabductive reasoning, mathematical arguments can undergo what we call anepistemic phase transition: a dramatic and rapidly-propagating jump fromuncertainty to near-complete confidence at reasonable levels of claim-to-claimerror rates. To show this, we analyze an unusual dataset of forty-eightmachine-aided proofs from the formalized reasoning system Coq, including majortheorems ranging from ancient to 21st Century mathematics, along with fourhand-constructed cases from Euclid, Apollonius, Spinoza, and Andrew Wiles. Ourresults bear both on recent work in the history and philosophy of mathematics,and on a question, basic to cognitive science, of how we form beliefs, andjustify them to others.