A study is made of the accuracy that could be obtained in the transfer function estimate. The system is noise-free but complex. It is found that it is better to use a (very) high-order model in combination with suitable regularization (prior). The reason is that the potential problem with an ill-conditioned regression matrix (high variance) is counteracted by the regularization. The bandpass filters cause information delay in the transient phase and degrade model-error attenuation in the stationary phase. It is also shown how set membership identification can handle model errors and provide hard (and in the example provided, somewhat conservative) error bounds. For this, the a priori information and the fact that the system is noise-free are used. An additive, bounded disturbance would not cause problems. The actual set membership estimate (the center of the model set) does not differ very much from the least-squares estimate in the example (because of the particular choice of prior).