Capitalism can be thought of as a complex learning system. We, as a society, want people to contribute the most that they can to the overall goals of humanity and capitalism acts as a vehicle for that.
Comparing it to another popular complex learning system, neural networks, might provide insight on failure conditions.
Income taxes can be thought of as gradient regularizers.
Basically limit the amount of weights accumulated per training session (wealth per tax year). Estate taxes can be thought of as weight regularizers (that occur once per lifetime). Maybe it would be more palatable as a tax on wealth yearly.
One of the ways that networks fail (and capitalism can fail) is with extreme values on weights (wealth) on small portions of the network. There are ways to counter this on a neural network - some are heavy handed (l1, l2 regularization), but some are just clever without direct analogies (dropout, batch normalization, etc).
I wonder how the conversation would change if we had to calculate our income tax as a percentage of total wealth each tax year.