Add Aggregate Function geomean (mosaic-sql)
#684
Conversation
| function geomeanExpr(preagg, node) { | ||
| const as = addStat(preagg, node); | ||
| const { expr, name } = countExpr(preagg, node); | ||
| return pow(product(pow(as, name)), div(1, expr)); |
There was a problem hiding this comment.
This looks "theoretically" correct to me: given a set per-bin geometric means, exponentiate by the (bin-level) count to "undo" the bin-level nth-root, multiply the results, then take the (global-level) nth-root.
However, might this suffer from overflow? That intermediate product could get very large. (The same result as if we just used product as the sufficient statistic rather than geomean, which would also simplify the overall scheme to just a single pow call for the nth-root.)
This might be worth testing more with large numbers. If we see issues, a more robust alternative would be to instead use the sum of log values as the sufficient statistic. Then the output aggregate expression would be exp(div(sum(as), expr)).
There was a problem hiding this comment.
Yes, I agree.
The code is already updated as it is quite clear, that there will be numerical issues.
Note: DuckDB uses log-based computation for geomean as well (see here).
This pull request introduces a new aggregate function —
geomean— to themosaic-sqlpackage for computing the geometric mean.What does this pull request cover?
geomeanaggregate functionsufficient-statistics.js)preaggregator.test.jsandaggregate.test.js)What is missing?
The JSON schema file has not yet been updated.
It is currently unclear whether this file will be auto-generated or if a manual modification is required.