Theta sketch and HLL numbers have more than 3% error

I have a couple of users column, one of them uses hyperUnique column type and second one uses thetaSketch column type during data ingestion.

The original value of distinct counts of users for a single day in Hive is 38030.

But when i ingest the HLL and thetaSketch columns in Druid, i get 35240 (HLL), 35530 (ThetaSketch).

The error in both cases is about 6.5 %

Does anyone know why the error could be so high? I even tried thetaSketch with size 131072 which is supposed to give 0.83% error in 99.73 percent cases.