When you do a calculation that involves the calculation of the Hartree-Fock exchange (e.g. calculations with hybrid functionals, including using the ADMM method), you may occasionally encounter the following warning message (more often if ADMM is used):

The Kohn Sham matrix is not 100% occupied. 
This may result in uncorrect Hartree-Fock results.
Try to decrease EPS_PGF_ORB and EPS_FILTER_MATRIX
in the QS section.

As you may already know, CP2K extensively uses the sparsity of matrices to increase efficiency of calculations. In a normal DFT (LSDA/GGA) calculation, the sparsity of the Kohn Sham (KS) matrix is the same as the overlap matrix, and therefore only parts of the density matrix that are non-zero in the corresponding parts of the overlap matrix are needed for the calculation. In other words, the density matrix and Kohn Sham matrix are stored using the same sparse pattern as the overlap matrix. This makes the calculation efficient by avoiding unnecessary multiplication with zero matrices blocks, and the accuracy of the calculation as well as the electron density function are unaffected.

For Hartree Fock exchange calculations, however, this is no longer the case. The sparsity of the KS matrix is no longer that of the overlap matrix. For the HF exchange term, we need all non-zero blocks of the density matrix and not only those that matched with the overlap. This contributes to more non-zero elements in the KS matrix.

Due to anything outside the sparsity pattern of the KS matrix—which is based on that of the overlap matrix—is not stored, the Hartree Fock code performs a screening of the HF exchange terms based on the sparsity pattern of KS, and has to assume that if a block is not present in the KS matrix then the HF exchange contribution to the block is also zero. This is not always true, depending on the property of the density matrix. Therefore, when the code detects that non-zero contributions of the HF exchange terms are being screened out, it produces the above warning message.

EPS_PGF_ORB controls the sparse pattern of the overlap matrix. Any element in the overlap matrix whose absolute value is smaller than EPS_PGF_ORB is treated as zero. And if a whole atomic block of the overlap matrix is less than EPS_PGF_ORB then the block will not be included in the sparse matrix.

The default value for EPS_PFG_ORB is set as the SQRT of the value of EPS_DEFAULT.

On top of KS matrix having sparse pattern of the overlap matrix, if EPS_FILTER_MATRIX is not zero, then a filtering process is performed so that any atomic blocks in the KS matrix having all elements with their absolute values less than EPS_FILTER_MATRIX is then removed from the KS matrix.

The default value of EPS_FILTER_MATRIX is zero, so this extra filter step is not used unless specified by the user.

If EPS_FILTER_MATRIX is already zero, then setting EPS_PGF_ORB to a smaller value will eventually remove the warning message. If EPS_FILTER_MATRIX is not zero, then the first course of action is to reduce it.

As usual with screening, the typical error you make is on the order of EPS, except the case when the calculation becomes unstable, which yields results that can be essentially unrelated to EPS (e.g. wrong by O(1)). In practice, despite of the warning, if a calculation is stable it should be accurate up to the value of EPS_PGF_ORB or EPS_FILTER_MATRIX—whichever is larger. You can always check by running a single point calculation with a smaller EPS value, and see if the difference in the total energy is in the order of magnitude as the larger of the EPS value you have tested.

The warning message is therefore more for the case when the calculation becomes unstable due to the forced screening of the HF exchange. So if your calculation is unstable, and if you see this warning message, then EPS_PGF_ORB and EPS_FILTER_MATRIX are the first places to look.

Note that by decreasing EPS_PGF_ORB, you will be making the overlap matrix more dense, and correspondingly also the density and KS matrices stored in CP2K. This will increase your computational cost. However, for calculations with hybrid functionals the cost of HF exchange term usually dominates and hence the associated cost increase due to more dense overlap, KS and density matrices may not be significant in comparison.