This keyword estimates the distinct key count for a partitioned index file instead of calculating the exact distinct key count by summing the distinct key counts for each partition of that index. This behavior can significantly reduce the time necessary to calculate the distinct key count for a partitioned index (at the cost of a larger uncertainty in the estimate).
Default sampling behavior is to only sample three active partitions unless PARTITION_ESTIMATE_LIMIT sets this limit to a value greater than three. The partitions sampled are the first and last partitions that ordinarily would have been used, and one or more in the "middle" of the remaining active partitions. At this time no attempt is made to spread out the sampled middle partitions over the range of available middle partitions.
With a negative value, the behavior resorts to summing the distinct key counts for each active partition.