Which VICReg term directly prevents complete collapse?

Assume both branches of a self-supervised VICReg model output the same constant vector for every input. Which objective component most directly penalises this solution?

The variance term, because each embedding dimension has insufficient standard deviation

The invariance term, because the two branches would be too similar

The covariance term alone, because constant dimensions have maximal off-diagonal covariance

The feature-normalisation operation, because constant embeddings cannot be normalised