Linear cryptanalysis is a well-known attack based on linear approximations, and is said to be feasiblefor an n-bit block cipher if the data complexity is at most 2. In this paper we consider IDEA with independent and uniformly distributed subkeys, referred to as IDEA with extended subkeys. We prove that any linear approximation of IDEA with extended subkeys, generalized to R rounds, requires at least approximations to the multiply operation. We argue that the best approximations are based on approximating least significant bits in the round operations and show that the probability of selecting a key for which such a linear cryptanalysis is feasible on IDEA is approximately 2.