Skip to main content
19 events
when toggle format what by license comment
Jan 29 at 18:42 comment added Ulrich Neumann @xzczd Thank you very much ! Our discussion was very inspiring. Your campaign concerning the extra-bounty is unprecedented and has touched me.
Jan 29 at 15:33 history bounty awarded xzczd
Jan 29 at 15:33 comment added xzczd This answer is definitely underestimated at the moment, so here's a small gift, +11 :) .
Jan 21 at 11:34 comment added Ulrich Neumann @xzczd Increasing the number of neighbors possibly decreases the approximation of boundary points
Jan 21 at 9:47 comment added Ulrich Neumann @xzcd Thanks for reply. I'm unsure about using weights because the unknown grid values can't be separated
Jan 21 at 9:21 comment added xzczd Today I revisited related papers e.g. sciencedirect.com/science/article/pii/S0307904X0300091X , two small discoveries: 1. A larger neighbor size (e.g. 12+1 instead of 8+1) seems to improve the result significantly. 2. Use a weight function (define Block[{d,dm}, w = Function[{d, dm}, #] &[With[{x = d/dm},If[d <= dm,1 - 6 x^2 + 8 x^3-3 x^4,0]]//PiecewiseExpand//Simplify`PWToUnitStep]] then in pinv use w[Sqrt[(xi - xj)^2 + (yi - yj)^2], dm] {xj - xi, yj - yi, …}) also improve the result, but the papers don't seem to show a general way to set dm.
Jan 18 at 14:42 comment added Ulrich Neumann I am looking forward to the answer
Jan 18 at 11:34 comment added xzczd This is surprising. I leaved a comment under that post, hopefully J.M. will finally have a look at this. (He seems to be busy these days… )
Jan 18 at 9:18 history edited Ulrich Neumann CC BY-SA 4.0
added 2 characters in body
Jan 18 at 8:59 comment added Ulrich Neumann @xzczd If of interest, here my testcode mat = RandomReal[{-1, 1}, {RandomInteger[{4, 20}], 8}]; rS = Table[u[i], {i, 1, Dimensions[mat][[1]]}]; erg1 = LeastSquares[mat, rS] // AbsoluteTiming; erg2 = PseudoInverse[mat] . rS // AbsoluteTiming; {Norm[erg1[[-1]] - erg2[[-1]]], erg1[[1]]/erg2[[1]]}
Jan 18 at 8:37 comment added Ulrich Neumann @xzczd Thanks for your interesting reply. I checked LeastSquares and PseudoInverse for different numerical examples, all with a symbolic righthandside. Both commands gave the same result (not surprising) but in all examples PseudoInverse evaluated significantly faster (surprising for me)
Jan 18 at 3:43 comment added xzczd Interesting. Perhaps I should revisit those papers. Two quick suggestions: 1. LeastSquares may be a better choice compared with PseudoInverse: mathematica.stackexchange.com/questions/117979/… 2. The AxesLabel can be modified to e.g. HoldForm@D[f, x, y] // TraditionalForm.
Jan 17 at 19:26 history edited Ulrich Neumann CC BY-SA 4.0
added 2771 characters in body
Jan 17 at 17:29 history edited Ulrich Neumann CC BY-SA 4.0
added 397 characters in body
Jan 17 at 16:53 history edited Ulrich Neumann CC BY-SA 4.0
added 51 characters in body
Jan 17 at 16:46 history edited Ulrich Neumann CC BY-SA 4.0
added 40 characters in body
Jan 17 at 16:41 history edited Ulrich Neumann CC BY-SA 4.0
added 40 characters in body
Jan 17 at 16:35 comment added Ulrich Neumann @xzczd See my new answer, perhaps it shows an easy access to the the GFDM method you mentioned today.
Jan 17 at 16:33 history answered Ulrich Neumann CC BY-SA 4.0