If the NDV becomes more than the numberBuckets, the histogram becomes very degenerate. E.g. for numberBuckets = 10 , all the distinct values beyond the first 9 that comprise almost all of the rows in the dataset are put together in the last bucket. This leads to very skewed estimates. The histogram is neither equal-width nor equal-depth and does not seem to minimize the estimation error in any discernible way.

Value N is present in the table N times:

db.foo.drop();
let docs = [];
for (let q = 0; q < 1000; q++) {
    for (let i = 0; i < q; i++) {
        docs.push({a: q});
    }
}
db.foo.insert(docs);
db.foo.runCommand({analyze: "foo", key: "a", numberBuckets: 5});

This produces the following bounds:

bounds: [ 1, 2, 4, 5, 999 ] 

All the values ~6 to ~999, present in 498486 documents get a single bucket. The rest of the histogram is occupied by values 1 to 5, present in 15 documents only. So we have the following wildly inaccurate estimates:

Enterprise test> db.foo.find({a:7}).count();
7
Enterprise test> db.foo.find({a:7}).explain().queryPlanner.winningPlan.cardinalityEstimate;
502