Hi Maribel

First of all, how much is a few and what sort of data you have? Also, what is your final purpose (what do you want to get)? 

PCA is mainly qualitative analysis. You can use it to try and find from what the signal is composed of. It can also be a pre-processing tool which is used to reduce the data to a much smaller and thus easier to analyze (although you can also lose data this way.

PLS is mainly for quantitative analysis. To retrieve a good estimation result you need a data set which is usually larger than a "few". 

Both methods are similar in such a way that they extract a number of latent variables from the set and reduce the data. The number of the samples depends on the complexity of it. For FTIR samples I could say that you need at least a few dozens, and more if you have several overlapping components.

If you want to use it for classification than you should use only PCA (you can use other methods such as NN and SVM etc.) and not PLS.

In order to try and help you should be more specific 

If its something simple and not heavy I can even run it for you using my codes for preliminary results which can guide you further.

Oz

one of my professor always said that you can have three clusters with three samples, but the question is how many variables you are considering for calculating your PCA. I agree that for FTIR probably it doesn't make sense to use less that a few dozens samples.

Hi Maribel,

Genrally you can't have more latent variables than the number of your samles. It is also important to know, the more replicate measurments you introduce to the data matrix, the better, more reliable your your model  will be. 

According to me you can use PLS-DA for qualitative analysis.As OZ said, it uses slightly different calculation than PCA, and sometimes gives better classification. You simply use numbers as indicator variables. You can't of course make any quantitative estimation then.

The question is still the same, what you are planning to use the model for? There a othe classification methods such as cluster analysis, CART, power tree etc. 

Success,

János