How does 'estimate' determine the first error terms?

When I estimate the parameters of an ARMA(1,1) model to fit on my time series using 'estimate', I don't understand how this function computes the first error term of the time series. To fit the parameter of the autoregressive part, I can imagine that the function just starts at t = 2, but how does the function know the error term at t = 2 when it does not know where to compare this observation to?