c.googleusercontent.com/external_content/untrusted_dlcp/research.google.com/en/us/pubs/archive/40671.pdf
Brief HyperLogLog++ Overview
Uses 64 bit hashing instead of 32 Has two representation modes: sparse and normal
'normal' is approximately the same as regular hyperloglog (still uses 64 bits)
'sparse' handles lower cardinality values with a highly accurate but poorly scaling strategy and leverages data compression to compete with 'normal' for as long as possible (sparse has the advantage on accuracy per unit of memory at low cardinality but quickly falls behind).