[omnicognate]

Oops, realised I got my footnote [2] the wrong way round here, and it's too late to edit it. In forward mode we need one evaluation for each input to get all the sensitivities. In reverse/adjoint mode it's one per output. In real world situations we're often performing calculations on a large set of input numbers to produce a small set of output numbers we're interested in, so there may be orders of magnitude fewer outputs than inputs.