This is a MATLAB implementation of the "Least Angle Regression" as in 
Bradley Efron, Trevor Hastie, Iain Johnstone, and Robert Tibshirani.
Ann. Statist.  32 (2004), no. 2, 407--499

Three examples are included in example[0-2].m. Example0 contains a
uniform random synthetic data for regression. Example1 
contains the diabetes example as is included in the reference paper.
Example2 tests tie breaker in case of two covariates having the
same correlation.

The function lars.m is the central function. It has line-by-line comments
for close study.

CAVEAT:
A few things to note about this implementation:
(1) It is optimized for efficiency and is less concerned 
    with clarity of the exposition of the algorithm.
(2) It does not implement the variants in the paper, i.e., Lasso or 
    forward stagewise selection.
(3) The covariate are suppposed to be linearly independent, otherwise
    an error message is reported.

CHANGELOG:

August 16 2004, major rewrite

special thanks to Quaid Morris for discussion aand suggestions!

(1) address numerical problems:
  . the effect of eps_tol on large dataset is offset by normalizing 
    covariate and response to std dev=1 
  
(2) handle boundary cases:
  . tie should be considered but otherwise proceed as usual
  . chose not to include collinear covariates to maintain simplicity

(3) speed up the code:
  . derived formulae in paper are transformed for efficiency
  . changed pinv to QR decomp and then to Cholesky decomp
  . speed up Cholesky by home brew update (same as Hastie's code)
  . change the gamma computation at last step
 
(4) logistics:
  . added max_steps
  . added tracep


April 18 2004,  initial coding and testing