There are several options to explore here, and its not clear of the advantages of each:

• Option 0: simple scaling factor
• Option 1: use set of samples to estimate correlation rather than simple scaling factor
• Option 2: exponential backoff
• Option 3: Use NDVs for each field as a weighting mechanism (not sure if this is the approach described by Milena or not)
• Option 4: Multi-field histograms (would require equality on prefix predicate?)

The 4th option is out of scope for this ticket but would be worth considering down the line.