Chazin Home | Ca-binding Protein DB | Vanderbilt Home
	Research Description | Publications | Wisdom | Search
	How to contribute | About this page

Linear prediction do's and don'ts

• Have all but one dimension Fourier transformed before you attempt linear prediction. The additional modulations introduced by untransformed dimensions cause stability problems for LP and will degrade the quality of the resulting spectrum.
• If you want to linear predict more than one dimension, use this scheme in a Felix macro:

for plane 1 N
   for fid 1 M
     rn datfile
     sb 512 90
     ft
     red
     sto 0 &row &plane
   next
next

bun 2
for vec 1 &vector
  lwb
  def datype 1
  ft     <--- NO WINDOW APPLIED
  red
  stb
next

bun 3
for vec 1 &vector
  lwb
  def datype 1
  lpx 1 N N+1 N+N/4 8 2 1
  sb N+N/4 90
  ft
  red
  stb
next

bun2
for vec 1 &vector
  lwb
  hft 1  <--- FLAG IS IMPORTANT
  ift    <--- CAN ONLY WORK, IF LAST POINT WAS NOT APODIZED TO 0
  lpx 1 M M+1 M+M/4 8 2 1
  sb M+M/4 90
  ft
  red
  stb
next
bun 0

• Use forward-backward prediction, which has been proven to be superior to the forward method (Zhu and Bax, JMR 100, 202-207, 1992), but is slower (<method>=2 for Felix lpx command).
• Test your parameters on a few vectors and fiddle by hand until you are sure what you want to do.
• With most spectra you can safely linear predict 30% more data. Remember that you will severely deemphasize the artificial data by windowing and hence will not feel mathematical problems. The beauty of this is that you won't deemphasize data that you spent time measuring!
• If your FID does not decay - in constant time or semi-constant time experiments or severely truncated data such as 4D - you should use the mirror image linear prediction (Zhu and Bax, JMR 90, 405-410, 1990). This provides more mathematical stability and lets you predict out to twice the data size (80% more points is probably safer).

see also:

lpx - General Linear Prediction

Felix Command Reference Manual