Highly energy and performance efficient embedded computing through approximately correct arithmetic: a mathematical foundation and preliminary experimental validation