Numerically-stable $\log(\exp(a) + \exp(b))$ via the max-shift trick.

Composite over maximum_op + sub_op + exp_op + add_op + log_op. Factoring out the per-pair max keeps the exponent's argument $\le 0$ so the underflow side bears the rounding error rather than the overflow side.

Math

See Also

logsumexp_op — generalised reduction over axes.