... or worse; it is still true for any a. You could easily get {1,000,001, -1,000,000}, which for perfectly clean, precise, representable data is equivalent, but which magnifies any noise/error in one of the predictors by a million. or a billion.