Instantaneous arithmetic computation via ratio-encoding in chemical reaction networks

What the math gives to ML

The paper's transferable asset is a gauge-invariant representation of numbers by nonnegative pairs: the value is a ratio, so common rescaling of both channels has no effect. Pairwise tensor and direct-sum operations compose exactly into multiplication and addition, while signed dual rails implement subtraction and signed multiplication without requiring negative concentrations. This suggests a neural arithmetic architecture in which activations and optionally weights are stored as ratio-encoded channels, with explicit gauge renormalization after every operation to control dynamic range. The main opportunity is improved numerical stability for deep multiplicative or rational networks, although the representation costs extra channels and should be tested in targeted settings rather than ordinary unconstrained MLPs.