Looks TooGoodToBeTrue: AnInformation-TheoreticAnalysisofHallucinations inGenerativeRestorationModels