• by keeping the initial index search relatively quick
• Ix1 small (depends on number of unique key values)
• Ix2 larger (depends on amount of repetition of keys)
• typically, b_Ix1 ≪ b_Ix2

• making Ix1 sparse, since Ix2 is guaranteed to be ordered
• in this case, b_Ix1 = ceil( b_Ix2 / c_i )
• if Ix1 becomes too large, add Ix3 and make Ix2 sparse
• if data file ordered on key, could make Ix3 sparse

Ultimately, reduce top-level of index hierarchy to one page.