Least Squares by Iteration Wfit.wpj Abstract This document is about a fitting evenly spaced data at distances ri = iR/N to a Fourier sine series at frequencies fm = m/R. Sine transform.doc Sinetran.htm 2 2 R NR 1 H ( fm ) rh ( r )sin 2 f r dr rh m i (ri )sin 2 f m ri (1) f m 0 fm NR i 0 2 2 1 N R 1 h(ri ) fH ( f )sin 2 fri df f m H f m sin 2 f m ri (2) ri 0 ri R m 0 The left hand parts of (1) and (2) can use any values of f and r, while the right hand parts are restricted to fm=m/R and ri = iR/N. Note that r never quite makes it to R in the sum. This is not the place to worry about the mathematics of this. For h constant as r 0 and to zero as r R, the right hand parts of (1) and (2) can be simply regarded as trap rule approximations to the left hand integrals. Be aware however that for large f and r, these approximations produce strange periodicities. The 3-d version of this can be useful for theoretical considerations H( f L , m, n ) N 3 3 1 h(ri , j , k ) L N /2 i , j , k N / 2 N /2 , m, n N / 2 h ri , j , k exp ˆjf H f , m, n , m, n exp ˆjf ri , j , k , m, n ri , j , k (3) (4) The N and L in (3) and (4) are related to but not equal to the R and N in (1) and (2). In the appropriate limits, the h and H in (3) and (4) are the same as those in (1) and (2). This will be used below. Introduction The function h is to be approximated by hA (r ) c(r ) c(r ')hA (| r r ' |)d 3r ' (5) The integral in (5) is a 3-d convoution integral. Direct evaluation of this is detailed in Convolution in 3d.doc. In this document c and hA are functions of |r| which reduces the integrals to one dimensional integrals as detailed in Sine transform.doc. In transform space ConvDetail.doc. In addition, we are using the grid fm=m/R and ri = iR/N to reduce the transform integrals to the sum forms in (1) and (2). The function c(r) is a short range function c(ri) = 0 for ri > r0 or i < i0. (6) H f C f C f H f or Hf C f (7) 1 H f or C f Hf (8) 1 C f Where c. The values of c(ri) for r < r0 are to be determined by minimizing. 3 L 2 N wi h ri , j , k hA ri , j , k , c M /2 i , j , k M / 2 2 (9) Or equivalently 2 RM 2 4 wi ri 2 h ri hA ri , c (10) N i 0 In general im > i0 so that the values of c are determined. Newton-Raphson minimization of 2 with respect to c requires accurate first derivatives of hA with respect to c Extremal.htm .doc. Derivatives of h.doc discusses finding the first derivatives with respect to k = c(rk). H f m ; h(ri ; ) 2 N fm sin 2 f m ri (11) k ri m 1 k Or h(ri , j , k ; ) k 1 L 3 H f N / 2 1 , m, n ; k , m , n N / 2 exp ˆj 2 f , m, n ri , j , k (12) First Derivative array Using the (9) form 3 hA ri , j , k , c 2 L M /2 2 wi h ri , j , k hA ri , j , k , c k k N i , j , k M / 2 Then use the (12) form for the partial of hA with respect to k 2 L M /2 1 2 wi h ri , j , k hA ri , j , k , c k N i , j , k M / 2 L (14) Evaluate the sum on r first to form 3 L H DW f N 3 M /2 i , j , k M / 2 3 (13) H f N / 2 1 k , m , n N / 2 wi h ri , j , k hA ri , j , k , c exp ˆj 2 f , m, n , m, n ; exp ˆj 2 f ri , j , k (15) Note that this sum is not over all N or R 2 M H DW f m wi r h ri hA ri , c sin 2 f mri (16) N fm i 0 So that 2 1 2 k L 3 N / 2 1 , m , n N / 2 H f , m, n k ; H Dw f (17) , m, n Noting that the functions depend on |f| alone. Switch to the spherical coordinate representation 2 1 N 1 H f m ; (18) 2 4 f m2 H Dw f m k k R m0 Then from Derivatives of h.doc , m, n ri , j , k 2 1 2 N 1 1 2 C f m ; 8 rk 2 f m H Dw f m sin 2 f m rk 2 k R rk m 0 1 C f m ; (19) This is the sin transform of 1 2 C f m ; H Dw f m (20) 2 1 C f ; m Second Derivative array Using the equation (12) form, an approximate second derivative array is 3 h ri , j , k h ri , j , k 2 2 L M /2 (21) 2 wi k m k m N i , j , k M / 2 Substituting (12) into (21) 2 2 L 1 2 k m N L 3 6 M /2 H f N /2 i , j , k M / 2 wi , m, n k , m , n N / 2 ', m ', n ' N / 2 (22) Replace wi by <w> and evaluate the sum over r first 2 2 L 1 2 w k m N L 3 6 3 1 2 w L , m , n N / 2 ', m ', n ' N / 2 H f N /2 , m , n N / 2 H f N /2 , m, n , m, n k ; H f k ', m ', n ' m ; H f , m, n ; H f ; ; ', m ', n ' ; m N 3 exp ˆj 2 f , ' m, m ' n, n ' (23) m Switch to spherical coordinates H f ; H f m ; 2 2 1 N R 1 2 w 4 fm2 m k m R m0 m k (24) Inserting the partial of H values w 2 2 2 k k ' R N R 1 4 f m0 2 m 1 2 C f m ; C f m ; 1 2 C f m ; C f m ; 2 2 1 C f m ; k 1 C f m ; k ' Then the partial of C values w 2 2 2 k k ' R N R 1 4 f m0 2 m 1 2 C f m ; 2rk sin 2 f m rk 1 2 C f m ; 2rk ' sin 2 f m rk ' (26) 2 2 fm fm 1 C fm ; 1 C fm ; Simplify w 2 2 2 k k ' R 2 1 2 C f ; m sin 2 f r sin 2 f r 16 rk rk ' m k' m k 2 m0 1 C f ; m N R 1 (25) (27) This is a decided maximum for k = k', thus the matrix is nearly orthogonal. N R 1 1 2 C f m ; w 2 2 16 rk2 k ,k ' 2 k k ' R m 0 1 C f ; m 2 (28) , m, n f ', m ', n ' r i, j ,k Change in alpha k k 0 1 2 N 1 1 2 C f m ; 8 rk 2 f m H Dw f m sin 2 f m rk 2 R rk m 0 1 C f m ; 16 rk2 Or k k 0 1 rk N 1 f m0 m 1 2 C f ; w m 2 R m 0 1 C f m ; N R 1 2 1 2 C f m ; H Dw f m sin 2 f m rk 2 1 C f ; m 1 2 C f ; m w 2 m 0 1 C f m ; N R 1 2 (30) (29)