The Low Frequency Instrument aboard the PLANCK satellite will employ pseudo-correlation radiometers, operating over three broad bands centred at 30, 44 ,and 70 GHz. The radiometer scheme is based on the simultaneous comparison of two input signals, one coming from the sky and the other coming from a reference blackbody at a stable cryogenic temperature (near 4K) as close as possible to the sky temperature (about 2.7K). This choice is made in order to minimize non-white instrumental noise, typically exhibiting a 1/f spectrum. Effects due to the residual offset are minimised with a gain modulation factor applied in software. Fluctuations of the reference signal, due to fluctuation in the cooling chain or to straylight radiation, can also produce a parasitic signal which would mimic a true sky fluctuation. The PLANCK scientific goal of a high precision imaging of the CMB anisotropy requires an accurate characterisation of each part constituting the chain by using tools of modellisation and experimental tests. In this work we describe the concept of the radiometric chain, its functioning and the main sources of systematic errors, showing how, only with a hard modelling effort, it is possible to characterise, reduce and then remove in the data processing those systematic effects that may in principle compromise the quality of the whole instrument response.