My take is:

• To fully normalize your opeartional database.
• To denormalize your data warehouses, for mainly performance reason. (the chance that data go out of sync is much lower, as most likely you only add data to data warehouse, but don't update.)

I am everyday dealing with a not normalized operational database, the biggest problem we facing is "out of sync'd data". From time to time, when my users complaining some data is wrong, I go check database, some of the data is stored in 4 or 5 different tables, and they simply have different values, which one is right? which one is wrong? That's a living hell. You have to do database design right at the beginning, don't expect to fix anything later, that is just not possible.