Is there theory as to why in GANs, training the generator and discriminator intermittently proves ineffective?

It seems it is a common intuition many have while building GANs, to pause training the generator to let the discriminator catch up and vice versa, hoping for convergence. But from what I’ve read, the consensus is that this is ineffective, which is disappointing.

Is there any theoretical understanding of why something that seems this “obvious” doesn’t work? Many of the sources I’m reading are from the earlier days of GANs, and I don’t know if this understanding has changed in recent years.