TPC: fix the weight and grid scaling of the spline derivatives in addCorrections (#15994)

* TPC: fix the weight scaling of the mixed spline derivative in addCorrections

TPCFastSpaceChargeCorrectionHelper::addCorrections scaled the 4th parameter of each spline knot (the mixed derivative d2S/dudv)
with the square of the weight: secMainScale * secMainScale for the main correction, ls * ls * ... for the added ones (ls is the
weight times the z fade-out factor, 1 at the knots). A spline is linear in its parameters, so scaling the correction by w scales all
of them by w.

The correction was therefore right at the knots but wrong between them whenever a weight != 1 was used, e.g. the derivative map
added with weight lumiScale in TPCScalerSpec::buildMap. The error is proportional to w - w^2. Measured with a production map and
the derivative map of run 568041 (random points, vs the weighted point-by-point sum of getCorrectionLocal and of the inverse
getters), for map + 0.13 x derivative: rms ~20 um in dx and dz, ~10 um without the first 10 pad rows; deviations above 100 um are
rare (0.4 % of the points) and lie mostly in the first pad rows. At a weight of -0.5 the rms is ~150 um (~60 um without the first
10 pad rows). With the fix the sum is exact to float precision (< 0.1 um) for weights from -1 to 1.5, inverse corrections
included; w = 1 is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* TPC: fix the chain-rule grid factors in addCorrections

A correction added on a different spline grid is sampled at the main grid's knots, and its derivatives must be converted with the
chain rule: with u = (y - y0) * yScale, d/du_main = (yScale_corr / yScale_main) d/du_corr, the same for v. addCorrections used the
inverse ratios, and applied the u factor to the dS/dv parameter and the v factor to dS/du (knot parameters {S, dS/dv, dS/du,
d2S/dudv}).

For identical grids (e.g. the derivative map) the factors are 1 and nothing changes. The M-shape map is created on a 4 x 4 grid in
TPCScalerSpec, with factors of 35-98 with respect to a production map: adding the M-shape map of run 544797 (up to 6 mm) to a map of
run 568041 gave errors of up to 89 cm in dx and up to the kMaxCorrection limit of 100 cm in dy (rms 4 cm and 7 cm), vs the
point-by-point sum. With the fix: rms 1-2 um, max 60 um (resampling of non-nested grids) inside the main map's grid. Outside of it
(|z| > ~1.37 x for production voxel maps) the merged map follows the edge and z fade-out of the main grid, as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit a36f08c)
(cherry picked from commit f4d548e)

Author: 2 people

Detectors/TPC/calibration/src/TPCFastSpaceChargeCorrectionHelper.cxx

@@ -1040,8 +1040,8 @@ void TPCFastSpaceChargeCorrectionHelper::mergeCorrections(
         constexpr int nKnotPar3d = nKnotPar1d * 3;
 
         { // scale the main correction
-
-          double parscale[4] = {mainScale, mainScale, mainScale, mainScale * mainScale};
+          // a spline is linear in its parameters: scaling the function scales all of them, the mixed derivative included
+          double parscale[4] = {mainScale, mainScale, mainScale, mainScale};
           for (int iknot = 0, ind = 0; iknot < spline.getNumberOfKnots(); iknot++) {
             for (int ipar = 0; ipar < nKnotPar1d; ++ipar) {
               for (int idim = 0; idim < 3; idim++, ind++) {
@@ -1075,10 +1075,11 @@ void TPCFastSpaceChargeCorrectionHelper::mergeCorrections(
           double scale = additionalCorrections[icorr].second;
           auto& linfo = corr.getRowInfo(row);
 
-          double scaleU = rowInfo.gridMeasured.getYscale() / linfo.gridMeasured.getYscale();
-          double scaleV = rowInfo.gridMeasured.getZscale() / linfo.gridMeasured.getZscale();
-          double scaleRealU = rowInfo.gridReal.getYscale() / linfo.gridReal.getYscale();
-          double scaleRealV = rowInfo.gridReal.getZscale() / linfo.gridReal.getZscale();
+          // chain rule for the derivatives: u = (y - y0) * yScale, v = (z - z0) * zScale => d/du_main = (du_corr / du_main) d/du_corr
+          double scaleU = linfo.gridMeasured.getYscale() / rowInfo.gridMeasured.getYscale();
+          double scaleV = linfo.gridMeasured.getZscale() / rowInfo.gridMeasured.getZscale();
+          double scaleRealU = linfo.gridReal.getYscale() / rowInfo.gridReal.getYscale();
+          double scaleRealV = linfo.gridReal.getZscale() / rowInfo.gridReal.getZscale();
 
           for (int iu = 0; iu < gridU.getNumberOfKnots(); iu++) {
             double u = gridU.getKnot(iu).u;
@@ -1094,7 +1095,8 @@ void TPCFastSpaceChargeCorrectionHelper::mergeCorrections(
                 float lu, lv, ls;
                 corr.convLocalToGrid(sector, row, y, z, lu, lv, ls);
                 ls *= scale;
-                double parscale[4] = {ls, ls * scaleU, ls * scaleV, ls * ls * scaleU * scaleV};
+                // knot parameters: {S, dS/dv, dS/du, d2S/dudv}
+                double parscale[4] = {ls, ls * scaleV, ls * scaleU, ls * scaleU * scaleV};
                 const auto& spl = corr.getSplineForRow(row);
                 spl.interpolateParametersAtU(corr.getCorrectionData(sector, row), lu, lv, P);
                 for (int ipar = 0, ind = 0; ipar < nKnotPar1d; ++ipar) {
@@ -1110,7 +1112,7 @@ void TPCFastSpaceChargeCorrectionHelper::mergeCorrections(
               float lu, lv, ls;
               corr.convRealLocalToGrid(sector, row, y, z, lu, lv, ls);
               ls *= scale;
-              double parscale[4] = {ls, ls * scaleRealU, ls * scaleRealV, ls * ls * scaleRealU * scaleRealV};
+              double parscale[4] = {ls, ls * scaleRealV, ls * scaleRealU, ls * scaleRealU * scaleRealV};
 
               { // inverse X correction
                 corr.getSplineInvXforRow(row).interpolateParametersAtU(corr.getCorrectionDataInvX(sector, row), lu, lv, P);