Problem: When using SPSS, the result is always "non-positive definite matrices". And that is expected, once there are more variables in the analysis than there are cases, the correlation matrix will have linear dependencies, and KMO and Bartlett would result in non-positive definite matrices. Observation: in SPSS, independent variables are placed in columns and dependent variables are placed in rows, according to the recommendations of the tutorial.

Attempts: I've already tried to perform the KMO and Bartlett tests using pieces of the spectra, like 600-1000 cm-1, 1000-1400cm-1 and 1400-1800 cm-1. However, the spectra differ form each other by just some specific bands. The samples are authentic and counterfeit packages, and they differ only slightly by their pigments. Then, the result is still "non-positive definite matrices".

Founds: I've found a paper - Gautam et. al (2021) - that suggests that a matrix composed of 50x3000 (column*row = cases*wavenumbers) is analyzed with PCA and that same matrix was tested by KMO and Bartlett tests. So, it seems that they used independent variables in rows, contrary to what SPSS suggests in its layout.

When I put my data on that way, I obtain reasonable results from KMO and Bartlett tests.

My question: Concerning FTIR data, can I really put independent variables (wavenumber) in rows and dependent variables in columns to superimpose the problem of non-positive definite matrices?

Reference: R. Gautam, R. Chauhan, R. Kumar, and V. Sharma, “PLS-DA and infrared spectroscopy based rapid and non-destructive discrimination of black ball and gel pen inks for forensic application,” Forensic Sci. Int. Reports, vol. 3, 2021, doi: 10.1016/j.fsir.2020.100162.