;+
;NAME: barrel_sp_fold_m1.pro
;DESCRIPTION: BARREL mid-level spectral folding routine
;             method1 = single-parameter analytic spectrum, single drm
;
;REQUIRED INPUTS:
;subspec   background subtracted count spectrum
;subspecerr    its error bars
;model     spectral model of electron spectrum (default is exponential)
;          1 = exponential
;          2 = monoenergetic
;drm       response matrix for correct payload altitude and chosen PID
;          of electrons
;phmean    energy channel centers (should have length = length of spectrum)
;          -- this is for the photons dimension
;phwidth   energy channel widths (should have length = length of spectrum)
;          -- this is for the photons dimension
;ctwidth   energy channel widths (should have length = length of spectrum)
;          -- this is for the counts dimension
;ctmean    energy channel centers (should have length = length of spectrum)
;          -- this is for the counts dimension
;usebins   subset of energy channels (count space) to actually use for fitting
;maxcycles Maximum number of times to try rescaling range for fit parameters
;
;OPTIONAL INPUTS:
;quiet     Don't make graphs + screen output
;verbose   show some debugging info as fits go along
;
;OUTPUTS:
;params            best fit parameters
;param_ranges      ranges on best fit parameters (1 sigma) (2x2 array)
;modvals           values of the fit function at the centers of the energy bins
;chisquare         chi-square (not reduced)
;dof               degrees of freedom associated with chisquare
;
;CALLS:
;barrel_sp_fitgrid1.pro
;
;STATUS: Tested with artificial data, 7/20/12
;
;TO BE ADDED:
;     Support for other spectral models
;
;
;REVISION HISTORY:
;Version 1.0 DMS 7/18/12 -- split out from barrel_folding as new middle layer;
;                           fixed "dof" to use "usebins" (at the same
;                           time as fixing "fitgrid1" to use "usebins")
;Errors fixed 7/20/12 -- definition of "tryspec" didn't include
;                       multiplication by "phwidth", meaning lousy
;                       starting point for fits -- which converged anyway.
;Version 2.3 DMS ;8/26/12 -- fixed logical error in rescaling
;                      algorithm in search for chisquare+1 contour.
;                      Removed redundant routine identifier from "message"
;Version 2.4 DMS 8/26/12 -- added support for model 2 (monoenergetic)
;Version 3.4 DMS 4/17/14 -- save elecmodel for best fit e- spectrum
;Version 3.6 DMS 2/12/15    Use dynamic generation of first-guess
;                           parameters based on true altitude
;-

pro  barrel_sp_fold_m1, subspec, subspecerr, model, drm, phebins, phmean, phwidth, ctwidth, ctmean, usebins, maxcycles,  $
         params, param_ranges, elecmodel, modvals, chisquare, dof, altitude, quiet=quiet, verbose=verbose

;Find good starting parameters:
if model EQ 1 then begin
    ;;This formula for approximate e-folding from a count ratio between
    ;;two bands is empirical from simulations. 
    barrel_sp_testratio_exp, altitude,efold,efold_ratios
    w1=where(ctmean GT 110. and ctmean LT 150.)
    w2=where(ctmean GT 200. and ctmean LT 250.)
    rat =  total(subspec[w2])/total(subspec[w1])
    if rat LT min(efold_ratios) then startpar = min(efold) else $
       if rat GT max(efold_ratios) then startpar = max(efold) else $
          startpar = interpol(efold,efold_ratios,rat)
    tryspec = (exp(-phmean/startpar)*phwidth) # drm
endif else if model EQ 2 then begin
    barrel_sp_testratio_mono, altitude,emono,emono_ratios
    w1=where(ctmean GT 110. and ctmean LT 150.)
    w2=where(ctmean GT 200. and ctmean LT 250.)
    rat =  total(subspec[w2])/total(subspec[w1])
    if rat LT min(emono_ratios) then startpar = min(emono) else $
       if rat GT max(emono_ratios) then startpar = max(emono) else $
          startpar = interpol(emono,emono_ratios,rat)
    tryspec = phmean*0.
    w = (where( abs(phmean-startpar) eq min(abs(phmean-startpar)) ))[0]
    tryspec[w]=phwidth[w]
    tryspec = (tryspec) # drm
endif  else message, 'Only exponential or monoenergetic spectrum is currently supported.'

;Find a starting normalization by scaling area of model and data
;(this will be the same procedure for every starting model):
startnorm = total( subspec[usebins]*ctwidth[usebins] ) / total( tryspec[usebins]*ctwidth[usebins] )

;Try a starting range around these trial values.  If the minimum 
;chi-square is not on the boundary, zoom in.  If it is, zoom out.
;In either case, recenter.

points = 10   ;always run a 21x21(x21) grid
scaling = [0.5,0.5]  ;[norm,par]: best values +/- 50%

if keyword_set(verbose) then $
    print,'iter#', 'startpar','startnorm','bestpar','bestnorm','scalepar','scalenorm','bestchi',$
          format='(a8,4a11,2a13,a10)'

;Iterate the fit, adjusting the scale dynamically:
for i=0, maxcycles-1 do begin

    barrel_sp_fitgrid1, subspec, subspecerr, model, drm, phmean, phwidth, usebins, startpar, $
         startnorm, points, scaling, bestpar, bestnorm, bestparn, bestnormn, modvals, $
         chiarray, bestchi, pararray, normarray

   ;;if best value is not on boundary, zoom in or finish.
   ;;Note that zooming in or out on scalingdrm doesn't do anything if
   ;;you aren't using two drms.

   if abs(bestnormn) NE points and scaling[0] GE 0.001 then scaling[0] /= 1.5
   if abs(bestparn)  NE points and scaling[1] GE 0.001 then scaling[1] /= 1.5

   ;;If scaling is now very fine, break.  Note that the last values of the
   ;;scaling parameters recorded here aren't really the last
   ;;values used, the last value used could be 2.5 times higher in one or more:
   if scaling[0] LT 0.001 and scaling[1] LT 0.001 then break

   if abs(bestnormn) EQ points then scaling[0] *= 2.0
   if abs(bestparn) EQ points then scaling[1] *= 2.0

   if keyword_set(verbose) then $
      print,i,startpar,startnorm,bestpar,bestnorm,scaling[1],scaling[0],bestchi,$
          format='(i8,4f11.3,2f13.6,f13.4)'

   startpar = bestpar
   startnorm = bestnorm

endfor

;If it never got to the finest scale, break with error:
if scaling[0] GT 0.001 or scaling[1] GT 0.001 then $
    print, 'WARNING: Fit failed to converge in maximum number of cycles.'

;Set most output variables (either 2 or 3 best-fit params depending on
;treatment of response matrices:
params = [bestnorm, bestpar]
chisquare = bestchi
dof = n_elements(usebins) - 2

;Only one thing left: the error on the parameters.  This requires more
;effort.  Here we will wander radially outwards until we find that the
;whole boundary has chisq > chimin
;Always center on the best value:
startpar = bestpar
startnorm = bestnorm
points = 10

;Create masks for the outer boundary of the chi-square space:
edges1 = intarr(2*points+1,2*points+1)
edges2 = intarr(2*points+1,2*points+1)
edges1[0,*] = 1
edges1[2*points,*] = 1
edges2[*,0] = 1
edges2[*,2*points] = 1

;Create initial values for error bar search:
scaling = [0.1, 0.1]   ;first guess
scaling0 = scaling
minscaling = scaling
goingup = [0,0]

for i=0, maxcycles-1 do begin

    barrel_sp_fitgrid1, subspec, subspecerr, model, drm, phmean, phwidth, usebins, startpar, $
         startnorm, points, scaling, bestpar, bestnorm, bestparn, bestnormn, modvals, $
         chiarray, bestchi, pararray, normarray

  ;;First see if the contour is completely closed:
  ;;Look for chisq < min_chisq + 1 on boundary:
  w1 = where(edges1 and (chiarray LE chisquare + 1.),nw1)
  w2 = where(edges2 and (chiarray LE chisquare + 1.),nw2)
  nw=[nw1,nw2]

  ;;If the boundary is entirely outside of the chi-square contour, zoom
  ;;in by a factor of 2, unless you had already zoomed out, in which 
  ;;case you've actually identified the right scale:
  if (total(nw) EQ 0) then begin
       if (total(goingup) EQ 2) then break
       for j=0,1 do if (not goingup[j]) then scaling[j] /= 2.0
       continue
  endif

  ;;If boundary not entirely clear, take each axis separately, and
  ;;expand or contract the scaling:
  for j=0,1 do begin
    if (nw[j] GT 0) then begin
      goingup[j] = 1
      scaling[j] *= 2.0
    endif else begin  
      if (not goingup[j]) then scaling[j] /= 2.0 
    endelse
  endfor

endfor

;Now that we've found the appropriate scaling (within a factor
;of 2 of the point where the last good fit appears on the boundary), 
;do one very fine map of chisquare space to find the error bars:

points = 40

barrel_sp_fitgrid1, subspec, subspecerr, model, drm, phmean, phwidth, usebins, startpar, $
         startnorm, points, scaling, bestpar, bestnorm, bestparn, bestnormn, modvals, $
         chiarray, bestchi, pararray, normarray

;Pick out the subset of points within the min(chisquare)+1. contour:
w = where(chiarray LT chisquare + 1., nw)
if nw EQ 0 then message, 'Failure in finding error bars.'

;This makes up the last needed output parameter: ranges of the parameters
param_ranges = [ [min(normarray[w]),max(normarray[w])],[min(pararray[w]),max(pararray[w])] ]

;Added in v3.4: electron spectrum model values
if model EQ 1 then begin	
       elecmodel= bestnorm*exp(-phmean/bestpar)
endif else if model EQ 2 then begin   
       ;just as done in barrel_sp_fitgrid1.pro!
       elecmodel = 0.*phmean
       w = (where( abs(phmean-bestpar) eq min(abs(phmean-bestpar))))[0]
       elecmodel[w] = bestnorm 
endif

end