The gtol parameter is a mess. Trying to do something "sophisticated" like in #172 has led to terrible impulse response functions:
Using the old linear gtol * r scaling is already much much better:
Looking at the 2d integrated heatmap:
"Tuning" (read: randomizing) the magic constants in the _gtol_error function gives this
This is most likely due to Jacobians going all over the place at the boundaries of _gtol_error, which means we either need any old $C^\infty$ function, or just go back to the linear scaling.
The
gtol
parameter is a mess. Trying to do something "sophisticated" like in #172 has led to terrible impulse response functions:Using the old linear
gtol * r
scaling is already much much better:Looking at the 2d integrated heatmap:
"Tuning" (read: randomizing) the magic constants in the
_gtol_error
function gives thisThis is most likely due to Jacobians going all over the place at the boundaries of
_gtol_error
, which means we either need any old $C^\infty$ function, or just go back to the linear scaling.