y contaminated by noise because of multiple reasons, such as

nt error, sample preparation error and sample variance. Noise

adds more difficulty of the accurate and correct estimation of a

Because the baseline of a spectrum is unknown and not in an

format and importantly contaminated by noise, estimating the

for a spectrum therefore needs special algorithms. Most baseline

n algorithms work on two parts, i.e., the error must be minimised

moothness of a baseline curve must be maximised. Therefore

very baseline estimation algorithm has to trade off between the

error) and the smoothness. A model which strengthens too much

elity will end up with a peak spectrum, in which almost no signal

etected. However, a model which strengthens too much on the

ess will end up with a peak spectrum with too many artifacts

o many false signals and missed signals. The success of accurate

scovery is thus highly related with the selection of a proper

estimation algorithm. This chapter has introduced several

estimation algorithms including the most classical ones, such as

aker-Henderson algorithm, the spline algorithm and the wavelet

and the advanced one, such as the Bayesian Whittaker-

on algorithm. One important issue of discovering signals from a

ata set with multiple replicates is the alignment of the discovered

oss replicates. This is because the signals in a spectrum will have

rtainties. First, a signal may occupy an interval on the horizontal

spectrum (the spectra analyser). Therefore identifying a unique

nalyser value for a signal needs a smoothing process. Second, a

ay have some variation across replicates regarding the spectra

values. An appropriate alignment of the discovered signals across

must be considered. This chapter has therefore introduced a

proach to align discovered peaks across replicates within the

Whittaker-Henderson algorithm.