From 0e8345863eabb9ada61fbe3c9c03f393682a2615 Mon Sep 17 00:00:00 2001 From: Aleksander Alekseev Date: Tue, 12 May 2026 15:09:36 +0300 Subject: [PATCH v1] Decipher the comment in tsrank.c --- src/backend/utils/adt/tsrank.c | 34 +++++++++++++++++++++++++++------- 1 file changed, 27 insertions(+), 7 deletions(-) diff --git a/src/backend/utils/adt/tsrank.c b/src/backend/utils/adt/tsrank.c index d35e5528d0a..383ad393971 100644 --- a/src/backend/utils/adt/tsrank.c +++ b/src/backend/utils/adt/tsrank.c @@ -336,13 +336,33 @@ calc_rank_or(const float *w, TSVector t, TSQuery q) jm = j; } } -/* - limit (sum(1/i^2),i=1,inf) = pi^2/6 - resj = sum(wi/i^2),i=1,noccurrence, - wi - should be sorted desc, - don't sort for now, just choose maximum weight. This should be corrected - Oleg Bartunov -*/ + + /* + * The ideal score for a term is the weighted harmonic sum: + * + * resj = sum(wi / i^2, i = 1..noccurrences) + * + * where wi is the weight of the i-th occurrence and weights are + * sorted in descending order so that the highest-weight + * occurrence gets the smallest divisor (i=1) and thus contributes + * the most. + * + * The result is divided by pi^2/6 ~= 1.64493406685, which is the + * limit of sum(1/i^2, i=1..inf). This normalizes the score to the + * [0, 1] range. + * + * As an approximation for efficiency, we skip the sort and + * instead only promote the single highest-weight occurrence to + * position i=1. This is done by taking the raw (unsorted) sum + * resj, subtracting the maximum weight's actual contribution + * wjm/(jm+1)^2, and adding back its corrected contribution + * wjm/1^2 = wjm: + * + * adjusted = resj - wjm/(jm+1)^2 + wjm = wjm + resj - + * wjm/(jm+1)^2 + * + * The remaining occurrences are left in their original order. + */ res = res + (wjm + resj - wjm / ((jm + 1) * (jm + 1))) / 1.64493406685; entry++; -- 2.43.0