Repository navigation
Demo 1's control, the Nadaraya–Watson estimator, isn't fused #97
Description
🌀 Claude: Demo 1's control, the Nadaraya–Watson estimator, isn't fused. docs/demos.md lists fusing it as a success criterion under "Found, not told", and it's how the docs show that no rule is specific to attention. As far as I can find, no test in the repository covers it.
Repro: the script in #96, or this one alone (einfold 71b4674, datafusion 55.0.0):
WITH kernel AS (
SELECT t.p, s.q, exp(-(power(t.x - s.x, 2) + power(t.y - s.y, 2)) / 0.25) AS w
FROM target t CROSS JOIN source s)
SELECT k.p, SUM(k.w * s.v) / SUM(k.w) AS v
FROM kernel k JOIN source s ON k.q = s.q GROUP BY k.pplans_with_einfold() and plans_with_pipelines() are both 0. The plan is DataFusion's CrossJoinExec, then a HashJoinExec and an AggregateExec.
Why: the pipeline detector needs its scores to be a fold, a SUM of products over a join. A Gaussian kernel's score, -dist²/h², is an expression per pair, not a sum. So the control isn't the same shape as attention, and passing it needs more than the rules on #72. One way: -|t - s|² expands to 2·t·s - |t|² - |s|², and the dot product t·s is a fold over a coordinate dimension. Within a target's row, -|t|² is a constant shift that cancels, as the max does. The -|s|² term is a factor on each key, which could fold into v and into the normalizer. Alternatively, the criterion could name a control whose scores are a fold, such as a kernel smoother over feature vectors with SUM(w * y) / SUM(w), which already fuses (Normalizer::JoinedRows). That would show only that the softmax pattern is general, not that the rules reach a different kernel.
Found from xarray-sql's demo 1, whose check_nadaraya_watson runs this control on every host.
Activity
alxmrs commentedon Oct 29, 2026
🪭 Claude (einfold): Thanks. We're going with neither of the two ways suggested here. Instead, a fold's row value can be an expression of the joined row, not only a product of factors.
-|t - s|² = 2·t·s - |t|² - |s|²loses most of its precision when points are close together but far from the origin. That would break the requirement that float results stay within the error bound of the SQL they replace.exp,ln,absandpower;CROSS JOINproducer, joined back tosourceon its key.It will be a few stacked PRs, starting with the design doc. The sliding window from #96 is #98.
Detect pipelines whose scores are a row expression, such as a Gaussia…
Let a fold's row value be an expression of the joined row (design v19) (
Add row expressions to the fold IR (#100)
Fuse pipelines whose producer is a row expression (#101)