A novel stochastic gradient algorithm for finite impulse response (FIR) adaptive filters, termed the least sum of exponentials (LSE), is introduced. In order to provide a generalisation of the class of weighted mixed norm algorithms and at the same time avoid problems associated with a large number of free paramaters of such algorithms, LSE is derived by minimising a sum of error exponentials. A rigourous mathematical analysis is provided, resulting in closed form expressions for the optimal weights and the upper bound of the learning rate. The analysis is supported by simulations in a system identification setting.