Nilearn non_parametric_inference: why is neg_log10_vfwe_pvals_img thresholded?

The function non_parametric_inference with threshold=None and tfce=False returns neg_log10_vfwe_pvals_img. This image appears to be thresholded, in the sense that its value are non-trivial only in some particular regions (where it is large enough).
For instance, if I use the documentation Second-level fMRI model: one sample test - Nilearn and adapt the call to non_parametric_inference as follows:

neg_log10_vfwe_pvals_img = non_parametric_inference(
    n_perm=500,  # 500 for the sake of time. Ideally, this should be 10,000.

(basically I removed the threshold argument and renamed the output)
Then if I plot neg_log10_vfwe_pvals_img I get spots with non-trivial values (around the expected active regions) and then 0 values elsewhere.

This suggest that the map is already thresholded at some default significance level: indeed, without a threshold, I would expect the map to be non-trivial everywhere, and only after choosing some alpha-level should I restrict the values to consider for inference.

I am missing something? If there such a hidden default threshold? Note that the function does not take an alpha argument.

I think that the reason is that we are considering corrected p-values, i.e. the voxels p-values are compared to the min p-value (or max t statistic) distribution across a bunch of permutations. In many regions of the brain, the corrected p-value is thus equal to 1.
Does that answer your question ?

I wasn’t aware of this limitation of the max stat procedure, but thanks for pointing it out. It makes perfect sense!