Selective overweighting of larger magnitudes during noisy numerical comparison

Humans are often required to compare average magnitudes in numerical data; for example, when comparing product prices on two rival consumer websites. However, the neural and computational mechanisms by which numbers are weighted, integrated and compared during categorical decisions are largely unknown1,2,3,4,5. Here, we show a systematic deviation from ‘optimality’ in both visual and auditory tasks requiring averaging of symbolic numbers. Participants comparing numbers drawn from two categories selectively overweighted larger numbers when making a decision, and larger numbers evoked disproportionately stronger decision-related neural signals over the parietal cortex. A representational similarity analysis6 showed that neural (dis)similarity in patterns of electroencephalogram activity reflected numerical distance, but that encoding of number in neural data was systematically distorted in a way predicted by the behavioural weighting profiles, with greater neural distance between adjacent larger numbers. Finally, using a simple computational model, we show that although it is suboptimal for a lossless observer, this selective overweighting policy paradoxically maximizes expected accuracy by making decisions more robust to noise arising during approximate numerical integration2. In other words, although selective overweighting discards decision information, it can be beneficial for limited-capacity agents engaging in rapid numerical averaging.