**  denotes quite substantial/important changes
*** denotes really big changes 

ISSUES
------

* Some error message translation files need updating.

1.5-5

* `gamm.setup' made an assumption about basis dimensions which was not 
  true for tensor products involving the "cc" basis. This is now fixed.

1.5-4

* smooth.construct.tensor.smooth.spec modified, so that 
  re-parameterization in terms of function values is only if it's 
  stable, and by default the parameters are function values with
  even spacing. Otherwise it was possible for tensor products of
  p-splines to fail.

1.5-3

* `gam' now attempts to coerce `data' to a data frame, if it is not 
  already a list or a data frame, provided that it is already an object 
  that model.frame can deal with. This is to support an undocumented 
  feature of versions prior to 1.5-2 that `data' could actually be 
  something other than a list or data frame. 

* An offset of type "array" could cause gam.fit3 to fail. fixed.

* `variable.summary' bug fixed, (it caused gam(y~1) to fail).   

1.5-2

* Several exported functions had no usage entries in the help files.
  Everything exported does now.

* `vis.gam' had a bunch of bugs (which could make it fail altogether) 
   as a result of trying to set default conditioning values from the gam 
   object model frame. `gam' and `gamm' now obtain summary statistics of 
   the predictor variables, stored in `var.summary' in the gam object, 
   which `vis.gam' now uses. As a result `vis.gam' `view' and `cond' 
   arguments should now contain original variable names, not model frame 
   term names.

* `data' argument of `gam' no longer stored in the `gam' object, by 
  default to save memory (can restore this --- see `gam.control').

* `summary.gam' failed under na.exclude. Fixed. 

* `mroot' failed on 1 by 1 matrices, Fixed.

1.5-1

* The stability of the fitting methods had become substantially greater 
  than the stability of the edf calculations after fitting. So it was 
  possible to fit very poor models, and then have non-sensical effective 
  degrees of freedom reported. This has been fixed by using much more stable 
  expressions for edf calculation, when outer iteration smoothness
  selection is empolyed. (Changes are to gam.fit3, gam.outer and a new
  routine gam.fit3.post.proc).

* edfs in version 1.5-0 were calculated using newton, rather than fisher 
  weights, in the matrix F=(X'WX+S)^{-1}X'WX, the diagonal of which gives 
  the edf array. The problem with this is that it is possible for X'WX 
  not to be +ve definite, and then degrees of freedom can be non-sensical.
  Fisher weights are always used now (although the original problem is 
  exceedingly hard to generate an example of).

* The summation convention code could be *very* memory intensive for cases 
  in which the matrix arguments of a smooth feature many repeated values. 
  Code now fixed to make much more efficient use of any repeated rows in  
  matrix arguments. This enables much larger signal regression problems to 
  be tackled. 

* Some help file fixes.

1.5-0

*** Efficient and general REML/ML smoothing selection implemented. 
    Smoothness selection criterion and numerical optimizer are now
    selected using arguments `method' and `optimizer' of `gam', and 
    `gam.method' has been removed.

*** To further enhance stability and efficiency, Fisher scoring is now 
    only used for canonical links, when it corresponds to full Newton. 
    With non-canonical links PIRLS is based on full Newton. 

** Derivative iteration as in Wood (2008) has been replaced by a direct
   implicit function method (which costs no more given Newton based 
   PIRLS).

** An option `select' has been added to `gam' to allow terms to be
   completely removed from a model by smoothness selection. 

** The shrinkage smoothers "cs" and "ts" have been modified 
   substantially. The Wood (2006, 4.1.6) proposal of adding a 
   small multiple of the identity matrix to the penalty matrix 
   is flawed in that it tends to corrupt small eigen values
   of the penalty matrix for large (dimension) penalty matrices. 
   It is much better to set the zero eigenvalues of the penalty matrix 
   to a small proportion of the smallest +ve eigenvalue, and to 
   use the matrix with the resulting eigen-decomposition as the 
   penalty. This is now done. Thanks to Roman Torgovitsky for 
   reporting the original problem.

** Tweedie family added (including `ldTweedie' function to evaluate 
   log Tweedie densities for powers in (1,2]).

* "ps" "cp" and "cc" smooths can now be supplied with 2 knots to be 
  treated as `endpoints' of the smooths (full set of knots can still 
  be supplied as before).

* The `newton' optimizer was dropping terms when their gradient was 
  below the convergence threshold (and allowing re-entry). This 
  promotes zig-zagging unless the terms are independent. Now only 
  drops terms if gradient and second derivative are very small 
  (so obective is really flat). 

* The adaptive smoothing "ad" class has been greatly simplified and 2D 
  penalty improved. Much faster as a result, and 2D adaptive actually 
  quite good.

* gam.fit3 now checks that the initial PIRLS step produces an 
  improvement in penalized deviance relative to a null model. If not 
  then step halving towards the null model parameter is employed.
  The null model is as close to constant predicted values as the model
  structure allows (it is estimated up front in estimate.gam, to save 
  computation).

* `gam.side' now takes account of whether the model has an intercept 
   (or the model model matrix column corresponding to an intercept is 
   in the column space of the model matrix of the parametric model 
   components).  

* The smoothing parameter array returned by `gam' now includes names
  for the smoothing parameters.

* s and te check that `id' is a single element.

* By default, partial residuals are no longer plotted for smooths with 
  `by' variables since they are usually meaningless here (they can 
  be re-instated by argument `by.resids').

* `min.sp' processing modified to work with `paraPen' argument to 
  `gam'.

* vcov.gam defaults to Bayesian covariance matrix.
  
* indexing error in `parametricPenalty' corrected.

* plot.gam modified so that page change behaviour is like plot.glm 

* negbin family upgraded to work with (RE)ML.

* `sp.vcov' function added to extract covariance matrix of log 
  smoothing parameters from (RE)ML based fits.

* `power' links now handled by default fitting methods (i.e. gam.fit3)

* `magic.post.proc' now expects weights, not sqrt(weights) as the 
   `w' argument (unless `w' is a matrix).

* p-values tweaked again, for slightly better performance with smooths of 
  several variables. Still not quite right.

* record of intial sp's is now carried in `smooth' objects in field 
  `sp'.

* ?linear.functional.term error fix.

* memory leak in magic.c:magic fixed --- all fixed smoothing 
  parameters lead to 2 arrays being left unfreed.

* various .Rd file fixes.

1.4-2

* Some minor .Rd file fixes

1.4-1

* `Predict.matrix2' was not in NAMESPACE: fixed.

* term specific offsets handled properly w.r.t. `by' variables in 
  `smoothCon' (a rather specialized topic!)

* minor doc bug fix for `smooth.construct'.

1.4-0

*** Model terms can now include linear functionals of smooths, by 
    supplying matrix arguments and matrix `by' variables to model 
    smooth terms. This allows, for example, a model to depend on the 
    integral of a smooth function, or its derivative, or for models to 
    depend on functional  predictors. See ?linear.functional.terms. Main 
    code changes are in `smoothCon' and `predMat'.

** Smooth terms can now be linked in order that they have the same 
   smoothing parameters (and, by default, bases). Linkage is specified
   using the `id' argument to `s' or `te'. Terms with the same `id' value
   will have the same smoothing parameter(s).

** `by' variables can now be factor variables. Also smooth terms with a 
   `by' variable are only subject to a sum-to-zero constraint if it
   is needed for identifiability.

** Argument `paraPen' of `gam' allows (multiple) penalization of 
   parametric model terms. This allows `gam' to fit any model that
   can be expressed as a penalized GLM.

** p-values returned by `summary.gam' now default to a Bayesian 
   approximation which gives (substantially) better frequentist behaviour 
   than the old method.

** The 2 standard error bands for smooths shown by `plot.gam' can now
   include the uncertainty about the overall mean, by default. Such
   intervals have better coverage probability (of their target of 
   inference) than intervals for centred smooths. Argument 
   `type="iterms"' to `predict.gam' will return such standard errors.

** An adaptive smoother class has been added, for smoothing with respect
   to one or two variables: invoked with `s(...,bs="ad",...). 

** `gamm' now supports nested smooth terms, and uses the same, constraint 
   absorbed, parameterization as `gam'.

** `s' and `te' terms accept an `sp' argument setting the term specific 
   smoothing parameters (and over-riding argument `sp' of `gam'). Ignored
   by `gamm'.

** Negative binomial handling changed. `negbin' family added: adapted from 
   MASS to work with gam outer iteration fitting. `gam.negbin' fitting 
   routine added in order to enable use of `negbin' with outer iteration.
   See ?negbin for details. MASS families no longer supported. 
   `nb.theta.mult' removed from `gam.control'.

** The Eilers and Marx style p-spline class is now one of the default 
   smoothing classes, rather than just being an example of how to set up
   a class in the help file. cyclic versions are also available.

** `smoothCon' now handles `by' variables and centering constraints 
   automatically, removing the need for smooth constructors to do so.
   `PredMat' handles `by' variables automatically. Users can over-ride 
   this behaviour when adding smooth classes, if needed - see 
   documentation. 

** The interface for adding user defined smooths has been simplified, 
   but this may mean that some user defined classes which worked before
   no longer work: see ?user.defined.smooths

* `smooth.construct' methods are now expected to set default values
   for the penalty order `p.order' and the basis dimension `bs.dim'
   if none are supplied. They should also sanity check supplied 
   values. Previously this was done by `s', but this put unhelpful
   restrictions on new smooth classes.

* `smooth.construct' now expects to recieve `data' and `knots' arguments
   with names corresponding exactly to `object$term'. In addition `data'
   should contain only what is required by `object$term' + a final column
   containing a `by' variable, if present. Predict.mat expects the same of 
   its `data' argument. wrapper functions smooth.construct2 and 
   Predict.mat2 will accept a data frame containing any number of variables 
   -- all that is required is that `object$terms' can be evaluated using 
   it. These functions handle repeat rows in matrix arguments efficiently.

* bug fix in `plot.gam' -- no longer requires to hit return if `select'
  used (ever).

* bug fix in `fixDependence' --- a completely dependent `X2' would not
  be detected, since the first element of R2 would be zero: used first
  element of R1 to set scale instead.

* `gam.fit' passes corrected n to `magic' so that `n' used in gcv/ubre
   does not include obs with zero prior weight. `gam.fit3' already doing 
   this...

* `magic' and `gam.fit3' now allow log smoothing parameters to be a 
  linear transformation of a smaller set of underlying smoothing parameters.

* `mgcv' based fitting has been removed as an option in `gam', as has 
  Pearson based GCV. In consequence `am' argument removed from 
  `gam.method' and `globit' removed from `gam.control'.

* `get.var' now coerces matrix values to numeric vectors, to facilitate 
   the handling of linear functionals of smooths.

* `gam.fit2' has been removed, since gam.fit3 is simply better.

* The default optimizer for the generalized case has been made slightly 
  more efficient (derivative free evaluation of GCV/AIC has been 
  improved). The upshot is that the default is now faster than performance
  iteration in almost all cases (while still being more reliable).

* the `absorb.cons' option has been removed from `gam.control'.

* `fix.family.link' and `fix.family.var' bug fix --- only return 
   family unmodified if all required derivative functions are present.

* `smoothCon' now returns a list of smooth objects to facilitate factor
  `by' variables.

* `smoothCon' makes smooth object labels more informative, if there are 
  `by' variables... this also makes default plots more informative.

* `plot.gam' indicates `by' variables in labels

* `gamm.setup' modified to call `gam.setup' for most of the setup, leaving 
  just the re-parameterization step to do. 

* `gamm' modified to allow constraint absorption (same as `gam')

* `gamm' bug fixed whereby "cc" smooths would get the wrong null space 
  dimension (effect was small, but noticeable, in practice e.g. Cairo 
  temperature example from chapter 6 of Wood, 2006, book).

* print methods now return first argument invisibly as they should.

* code for (very) old style summary removed. 

* `gam.fit3' now traps derivative iteration divergence, and suggests 
  tightening the convergence tolerance `epsilon' in `gam.control'. 
  Divergence can happen for ill-conditioned models if the PIRLS has 
  not converged sufficiently.

* gamm.Rd updated to reflect change to gammPQL in 1.3-28.

1.3-31

* There was a most annoying warning generated by R 2.7.0 every time `gam'
  was used. Now there isn't.

1.3-30

* change to DESCRIPTION file.

1.3-29

* `magic' could segfault if supplied with many constraints and relatively 
   high rank penalties, so that after constriant the penalty  matrix 
   square roots had more columns than rows (never happened in additive 
   model case, but can happen in more general settings). Fixed.

* `gamm' now silently drops grouping factors within the correlation 
   structure formulae that duplicate random effects grouping factors 
   (which automatically act as grouping factors on the correlation 
   structures anyway).

* Some replacement of dubious `as.matrix' calls with use of `,drop=FALSE]' 
  in gamm.r   

1.3-28

** `gamm' modified to call a routine `gammPQL' in place of MASS::glmmPQL. 
  This avoids some duplication, and facilitates maintainance. 

* Bug fix in `formXtViX' where matrix dimensions got dropped when 
  subsetting thereby messing up variance calculations for gamm fits in 
  which some group sizes were 1. 

1.3-27

** Fix of nasty bug in large dataset handling with "tp" basis (introduced 
   in 1.3-26). Subsampling code was re-seeding RNG instead of intended 
   behaviour of saving RNG state and  restoring it. Fixed and tested.

1.3-26

* modification to `gam' so that GCV/UBRE scores reported with all fixed 
  smoothing parameters are consistent with equivalent under s.p. 
  estimation.

* gam.fit3 modified to test for convergence of coefficients as well
  as penalized deviance, otherwise in extreme cases the derivative 
  iterations can diverge.

* modifications of gam.setup, predict.gam and plot.gam to allow smooths
  to contribute an offset term to the model (offset is returned from 
  smooth.construct or Predict.matrix as an "offset" attribute of 
  model/prediction matrix). This is useful for smooths which have known 
  boundary conditions of some sort.

* PredictMat can now handle NAs in a returned prediction matrix.
 
* vis.gam can handle NA's in predictions.

** Modification of large dataset handling for "tp" and "ts" bases. If 
   there are more that 3000 unique covariate combinations for a tprs then 
   3000 combinations are randomly sub-sampled, and used as the initial 
   knots for tprs basis construction. The same random number seed is used 
   every time,  (R's RNG state is unaltered by this). Control of this is 
   usually via the `max.knots' (default 3000) and `seed' (default 1) 
   elements of the `xt' argument of `s'. In consequence, `max.tprs.knots' 
   has been removed from `gam.control'.

* Modification of `s' and `te' to allow an extra argument `xt' which can 
  contain extra information to pass to the basis constructors for smooths.

* removal of `full.call' from smooth.spec objects - it wasn't used 
  anywhere any more, and is a pain to maintain.

* removal of `full.formula' from the `gam' object - it is no longer used 
  anywhere and requires alot of code to construct.

1.3-25

* A bug in `null.space.dimension' caused prediction to fail for `s' terms 
  of 4 or more variables, unless the `m' argument was supplied explicitly 
  (and was large enough for the number of variables). Fixed. 

1.3-24

* summary.gam modified so that it behaves correctly if fitting routines 
  detect and deal with rank deficiency in parameteric part of a model.

* spring cleaning of help files.

* gam.check modified to report more useful convergence diagnostics.

** `model.matrix.gam' added.

** "cr" basis constructor modified to use the same centering conditions 
  as other bases (sum to zero over covariates, rather than parameters 
  sum to zero). This makes centred confidence intervals for smooths, of 
  the sort used in plot.gam, behave in a similar way for all bases. With 
  the old "cr" centering constraint there could be high negative 
  correlation between coefficients of a centered smooth and the intercept: 
  this could make centred "cr" smooth CIs wider than CIs for other bases 
  (not really wrong, but disconcerting).  

1.3-23

* step size correction bug fixed in gam.fit3. `Perfect' convergence could
  cause the divergence control loop to fail: the divergence control loop
  was asking for near strict decrease in the penalized deviance, which 
  could be numerically impossible to achieve if the algorithm had actually 
  converged completely.... fixed.  

* minor doc bug fixes.


1.3-22

* Cheap but unneccesary code added to gdi.c and magic.c to stop 
  inappropriate uninitialized variable warnings from some compilers.

** Bad bug in gam.fit3 fixed. Prior weights of zero were not handled 
  correctly - prior weight vector should have been subsetted before
  gdi call, but this didn't happen. Result was infinite derivatives
  and fit failure. fixed.

* Related bug in gam.fit3: dropped observations not handled correctly 
  in deviance calculation, which can result in inappropriate step 
  halving. fixed.

* inner loop 3 in gam.fit2 and gam.fit3 modified so that step halving 
  continues until penalized deviance is at worst non-increasing. 

* stupid bug in summary.gam, p-value calc. fixed.

1.3-21

* minor bug in gam.fit() - edf array not passed to `mgcv.find.theta'
  if method "perf.magic" used - so wrong EDF used for theta estimation 
  with neagative binomial. fixed. 

* Theta estimate added to family object of fitted gam if negative binomial 
  used...

* extract.lme.cov(2) modified to allow use with single level grouping 
  factors (not really sure when this is useful)

* bug in gam4objective called when using gam.method(outer="nlm") - never 
  used GCV.

* fixed bug in `newton' whereby immediate convergence actually caused
  routine to fail.

* modified `smoothCon' and `predictMat' so that `qrc' attribute always
  created if constraint absorption used, even if there are no constraints.
  This attribute can then be used to test that there are no unabsorbed 
  constraints (e.g. in `gam.outer').

1.3-20

* Bad bug in `newton' - step halving set up so that step *never* 
  accepted (it still beat all previous methods in simulations)

* Minor bug in `newton' step limiting of Newton steps reduced step
  to max component 1, rather than `maxNstep'. 

* Some documentation fixes

1.3-19

*** SUBSTANTIAL CHANGE: Improved outer iteration added via gdi.c coupled 
  with gam.fit3. Exact first and second derivatives of GACV/GCV/UBRE/AIC 
  are now available via new iteration methods. These improve the 
  speed and reliability of fitting in the *generalized* additive model 
  case. 

* numerous changes to NAMESPACE and gamm related functions to pass
  codetools checks.

** gam.method()  modified to allow GACV as an option for outer GCV 
  model selection.

* magic.c::mgcv_mmult modified so that all inner loop calculations are 
  optimal (i.e. inner loop pointers increments are all 1).

* `smooth.construct' functions for "cc" and "cr" smooths now increase `k'
  to the minimum possible value (and warn), if it's too low. 

** `gam' modified to allow passing of `mustart' etc to gam.fit and 
  gam.fit2, properly

* `gam' modified to fix a bug whereby fitting in two steps using argument 
  `G' could fail when some sp's are to be estimated and some fixed.

** an argument `in.out' added to `gam' to allow user initialization of 
  smoothing parameters when using `outer' iteration in the generalized 
  case. This can speed up analyses that rely on several refits of the same 
  model. 

1.3-18

* gamm modifed so that weights dealt with properly if lme type varFunc 
  used. This is only possible in the non-generalized case, as gamm.Rd 
  now clarifies.

* slight modification to s() to add `width.cutoff=500' to `deparse'

* by variables not handled properly in p-spline example in 
  smooth.construct.Rd - fixed.

* bug fix in summary.gam.Rd example (pmax -> pmin)

* color example added to plot.gam.Rd

* bug fix in `smooth.construct.tensor.smooth.spec' - class "cyclic.smooth"
  marginals no longer re-parameterized.

* `te' documentation modified to mention that marginal reparameterization 
  can destabilize tensor products. 

1.3-17

* print.summary.gam prints estimated ranks more prettily (thanks Martin 
  Maechler)

** `fix.family.link' can now handle the `cauchit' link, and also appends a
   third derivative of link function to the family (not yet used).

* `fix.family.var' now adds a second derivative of the link function to 
   the family (not yet used).

** `magic' modified to (i) accept an argument `rss.extra' which is added 
  to the  RSS(squared norm) term in the GCV/UBRE or scale calculation; (ii)
  accept argument `n.score' (defaults to number of data), the number to 
  use in place of the number of data in the GCV/UBRE calculation.
  These are useful for dealing with very large data sets using 
  pseudo-model approaches.

* `trans' and `shift' arguments added to `plot.gam': allows, e.g. single
   smooth models to be easily plotted on uncentred response scale.

* Some .Rd bug fixes. 

** Addition of choose.k.Rd helpfile, including example code for diagnosing 
   overly restrictive choice of smoothing basis dimension `k'.

1.3-16

* bug fix in predict.gam documentation + example of how to predict from a 
  `gam' outside `R'.

1.3-15

* chol(A,pivot=TRUE) now (R 2.3.0) generates a warning if `A' is not +ve 
  definite. `mroot' modified to supress this (since it only calls 
  `chol(A,pivot=TRUE)' because `A' is usually +ve semi-definite). 

1.3-14

* mat.c:mgcv_symeig modified to allow selection of the LAPACK routine
  actually used: dsyevd is the routine used previously, and seems very 
  reliable. dsyevr is the faster, smaller more modern version, which it
  seems possible to break... rest of code still calls dsyevd.

* Symbol registration added (thanks largely to Brian Ripley). Version 
  depends on R >= 2.3.0

1.3-13

* some doc changes

** The p-values for smooth terms had too low power sometimes. Modified 
  testing  procedure so that testing rank is at most 
  ceiling(2*edf.for.term). This gives quite close to uniform p-value 
  distributions when the null is true, in simulations, without excessive 
  inflation of the p-values, relative to parametetric equivalents when 
  it is not. Still not really satisfactory.

1.3-12

* vis.gam could fail if the original model formula contained functions of 
  covariates, since vis.gam calls predict.gam with a newdata argument 
  based on the *model frame* of the model object. predict.gam now 
  recognises that this has happened and doesn't fail if newdata is a model 
  frame which contains, e.g. log(x) rather than x itself. offset handling 
  simplified as a result.

* prediction from te smooths could fail because of a bug in handling the 
  list of re-parameterization matrices for 1-D terms in 
  Predict.matrix.tensor.smooth. Fixed. (tensor product docs also updated)

* gamm did not handle s(...,fx=TRUE) terms properly, due to several 
  failures to count s(...,fx=FALSE) terms properly if there were fixed 
  terms present. Now fixed.

* In the gaussian additive mixed model case `gamm' now allows "ML" or 
  "REML" to be selected (and is slightly more self consistent in 
  handling the results of the two alternatives).

1.3-11

* added package doc file

* added French error message support (thanks to Philippe Grosjean), and 
error message quotation characters (thanks to Brian Ripley.)

1.3-10

* a `constant' attribute has been added to the object returned by
  predict.gam(...,type="terms"), although what is returned is still not an 
  exact match to what `predict.lm' would do. 

** na.action handling made closer to glm/lm functions. In particular,
  default for predict.gam is now to pad predictions with NA's as opposed
  to dropping rows of newdata containing NA's. 

* interpret.gam had a bug caused by a glitch in the terms.object 
  documentation (R <=2.2.0). Formulae such as y ~ a + b:a + s(x) could 
  cause failure. This was because attr(tf,"specials") is documented as 
  returning indices of specials in `terms'. It doesn't, it indexes 
  specials in the variables dimension of the attr(tf,"factors") table: 
  latter now used to translate.

* `by' variable use could fail unreasonably if a `by' variable was not of 
  mode `numeric': now coerced to numeric at appropriate times in smooth
  constructors. 

1.3-9

* constants multiplying TPRS basis functions were `unconventional' for d 
  odd in function eta() in tprs.c. The constants are immaterial if you are 
  using gam, gamm etc, but matter if you are trying to get out the 
  explicit representation of a TPRS term yourself (e.g. to differentiate 
  a smooth exactly).

1.3-8

* get.var() now checks that result is numeric or factor (avoids 
  occasional problems with variable names that are functions - e.g `t')

* fix.family.var and fix.family.link now pass through unaltered any family 
  already containing the extra derivative functions. Usually, to make a 
  family work with gam.fit2 it is only necessary to add a dvar function.

* defaults modified so that when using outer iteration, several performance
  iteration steps are now used for initialization of smoothing parameters 
  etc. The number is controlled by gam.control(outerPIsteps). This tends
  to lead to better starting values, especially with binary data. gam, 
  gam.fit and gam.control are modified.

* initial.sp modified to allow a more expensive intialization method, but
  this is not currently used by gam.

* minor documentation changes (e.g. removal of full stops from titles)

1.3-7

* change to `pcls' example to account for model matrix rescaling changing 
smoothing parameter sizes.

* `gamm' `control' argument set to use "L-BFGS-B" method if `lme' is using 
`optim' (only does this if `nlminb' not present). Consequently `mgcv' now 
depends on nlme_3.1-64 or above.

* improvement of the algorithm in `initial.sp'. Previously it was possible 
for very low rank smoothers (e.g. k=3) to cause the initialization to 
fail, because of poor handling of unpenalized parameters. 

1.3-6

* pdIdnot class changed so that parameters are variances not standard 
deviations - this makes for greater consistency with pdTens class, and 
means that limits on notLog2 parameterization should mean the same thing 
for both classes. 

** niterEM set to 0 in lme calls. This is because EM steps in lme are not
 set up to deal properly with user defined pdMat classes (latter 
 confirmed by DB).

1.3-5

** Improvements to anova and summary functions by Henric Nilsson 
  incorporated. Functions are now closer to glm equivalents, and 
  printing is more informative. See ?anova.gam and ?summary.gam.

* nlme 3.1-62 changed the optimizer underlying lme, so that indefintie 
  likelihoods cause problems. See ?logExp2 for the workaround.
  - niterEM now reset to 25, since parameterization prevents parameters 
  wandering to +/- infinity (this is important as starting values for 
  Newton steps are now more critical, since reparameterization 
  introduces new local minima).

** smoothCon modified to rescale penalty coefficient matrices to have 
  similar `size' to X'X for each term. This is to try and ensure that 
  gamm is reasonably scale invariant in its behaviour, given the 
  logExp2 re-parameterization.

* magic dropped dimensions of an array inapproporiately - fixed.

* gam now checks that model does not have more coefficients than data.

1.3-4

* inst/CITATION file added. Some .Rd fixes

30/6/2005 1.3-3

* te() smooths were not always estimated correctly by gamm(): invariance 
  lost and different results to equivalent s() smooths. The problem seems
  to lie in a sensitivity of lme() estimation to the absolute size of the 
  `S' attribute matrices of a pdTens class pdMat object: the problem did 
  not occur at the last revision of the pdTens class, and there are no 
  changes logged for nlme that could have caused it, so I guess it's down
  to a change in something that lme calls in the base distribution. 
  To avoid the problem, smooth.construct.tensor.smooth.spec has been 
  modified to scale all marginal penalty matrices so that they have 
  largest singular value 1.

* Changes to GLMs in R 2.1.1 mean that if the response is an array, gam 
  could fail, due to failure of terms like w * X when w is and array 
  rather than a vector. Code modified accordingly.

* Outer iteration now suppresses some warnings, until the final fitted
  model is obtained, in order to avoid printing warnings that actually
  don't apply to the final fit.

* Version number reporting made (hopefully) more robust.

* pdconstruct.pdTens removed absolute lower limit on coef - replaced with
  relative lower limit.

* moved tensor product constraint construction to BEFORE by variable
  stuff in smooth.construct.tensor.smooth.spec.

1.3-1

* vcov had been left out of namespace - fixed.

* cr and cc smooths now trap the case in which the incorrect number of 
  knots are supplied to them.

* `s(.)' in a formula could cause a segfault, it get's trapped now, 
  hopefully it will be handled nicely at some point in the future. Thanks 
  Martin Maechler.

* wrong n reported in summary.gam() in the generalized case - fixed. 
  Thanks YK Chau. 

1.3-0

*** The GCV/UBRE score used in the generalized case when fitting by 
  outer iteration (the default) in version 1.2 was based on the Pearson 
  statistic. It is prone to serious undersmoothing, particularly of binary 
  data. The default is now to use a GCV/UBRE score based on the deviance: 
  this performs much better, while still maintaining the enhanced 
  numerical convergence performance of outer iteration.

* The Pearson based scores are still available as an option (see 
  ?gam.method)

* For the known scale parameter case the default UBRE score is now 
  just a linearly rescaled AIC criterion. 

1.2-6

* Two bugs in smooth.sconstruct.tensor.smooth.spec: (i) incorrect 
  testing of class of smooth before re-parameterizing, so that cr smooths 
  were re-parameterized, when there is no need to; (ii) knots used in 
  re-parameterization were based on quantiles of the relevant marginal 
  covariate, which meant that repeated knots could be generated: now uses 
  quantiles of unique covariate values.

* Thanks to Henric Nilsson a bug in the documentation of magic.post.proc has 
  been fixed. 

1.2-5

** Bug fix in gam.fit2: prior weights not subsetted for non-informative 
  data in GCV/UBRE calculation. Also plot.gam modified to allow for 
  consequent NA working residuals. Thanks to B. Stollenwerk for reporting 
  this bug.

** vcov.gam written by Henric Nilsson included... see ?vcov.gam

* Some minor documentation fixes.

* Some tweaking of tolerances for outer iteration (was too lax).

** Modification of the way predict.gam picks up variables. 
  (complication is that it should behave like other predict functions, but 
  warn if an incomplete prediction data frame is supplied -since latter 
  violates what white book says). 

1.2-2

*** An alternative approach to GCV/UBRE optimization in the 
  *generalized* additive model case has been implemented. It leads to more 
  reliable convergence for models with concurvity problems, but is slower 
  than the old default `performance iteration'. Basically the GAM IRLS 
  process is iterated to convergence for each trial set of smoothing 
  parameters, and the derivatives of the GCV/UBRE score w.r.t. smoothing 
  parameters are calculated explicitly as part of the IRLS iteration. This 
  means that the GCV/UBRE optimization is now `outer' to the IRLS 
  iteration, rather than being performed on each working model of the IRLS 
  iteration. The faster `performance iteration' is still available as an 
  option. As a side effect, when using outer iteration, it is not possible 
  to find smoothing parameters that marginally improve on the GCV/UBRE 
  scores of the estimated ones by hand tuning: this improves the logical 
  self consistency of using GCV/UBRE scores for model selection purposes.

* To facilitate the expanded list of fitting methods, `gam' now has a 
  `method' argument requiring a 3 item list, specifying which method to 
  use for additive models, which for generalized additive models and if using 
  outer iteration, which optimization routine to use. See ?gam.method for 
  details. `gam.control' has also been modified accordingly.
    
*** By default all smoothing bases are now automatically 
  re-parameterized to absorb centering constraints on smooths into the 
  basis. This makes everything more modular, and is usually user 
  transparent. See ?gam.control to get the old behaviour.
    
** Tensor product smooths (te) now use a reparameterization of the 
  marginal smoothing bases, which ensures that the penalties of a tensor 
  product smooth retain the interpretation, in terms of function shape, of 
  the marginal penalties from which they are induced. In practice this 
  almost always improves MSE performance (at least for smooth underlying 
  functions.) See ?te to turn this off.
    
*** P-values reported by anova.gam and summary.gam are now based on 
  strictly frequentist calculations. This means that they are much better 
  justified theoretically, and are interpretable as ordinary frequentist 
  p-values. They are still conditional on smoothing parameters, however, 
  and are hence underestimates when smoothing parameters have been 
  estimated.

** Identifiability side conditions modified to work with all smooths 
  (including user defined). Now works by identifying possible dependencies 
  symbolically, but dealing with the resulting degeneracies numerically. 
  This allows full ANOVA decompositions of functions using tensor product 
  smooths, for example.

* summary.gam modified to deal with prior weights in adjusted r^2 
  calculation.
    
** `gam' object now contains `Ve' the frequentist covariance matrix of 
  the paremeter estimators, which is useful for p-value calculation. see 
  ?gamObject and ?magic.post.proc for details.

* Now depends on R >=2.0.0
    
* Default residual plots modified in `gam.check'
    
** Added `cooks.distance.gam' function.
    
* Bug whereby te smooths ignored `by' variables is now fixed. 

1.1-6

* Smoothing parameter initialization method changed in magic, to allow 
  better initialization of te() terms. This affects default gam fits.
    
* gamm and extract.lme.cov2 modified to work correctly when the 
  correlation structure applies to a finer grouping than the random 
  effects. (Example of this added to gamm help file)
    
* modifications of pdTens class. pdFactor.pdTens now returns a vector, 
  not a matrix in accordance with documentation (in nlme 3.1-52). Factors 
  are now always of form A=B'B (previously, could be A=BB') in accordance 
  with documentation (nlme 3.1-52). pdConstruct.pdTens now tests whether 
  initializing matrix is proportional to r.e. cov matrix or its inverse 
  and initializes appropriately. gamm fitting with te() class tested 
  extensively with modifications and nlme 3.1-52, and lme fits with pdTens 
  class tested against equivalent fits made using re-parameterization and 
  pdIdent class. In particular for gamm testing : model fits with single 
  argument te() terms now match their equivalent models using s() terms; 
  models fitted using gam() and gamm() match if gam() is called with the 
  gamm() estimated smoothing parameters.
   
* modifications of gamm() for compatibility with nlme 3.1-52: in 
  particular a work around to allow everything to work correctly with a 
  constructed formula object in lme call.
  
* some modifications of plot.gam to allow greater control of 
  appearance of plots of smooths of 2 variables.
  
* added argument `offset' to gam for further compatibility with 
  glm/lm.
  
* change to safe prediction for parameteric terms had a bug in offset 
  handling (offset not picked up if no newdata supplied, since model frame 
  not created in this case). Fixed. (thanks to Jim Young for this) 1.1-5
    
* predict.gam had a further bug introduced with parametric safe 
  prediction. Fixed by using a formula only containing the actual variable 
  names when collecting data for prediction (i.e. no terms like 
  `offset(x)') 

1.1-5

* partial argument matching made col.shade be matched by col passed in 
..in plot.gam, taking away user control of colors. 1.1-5
    
* 2d smooth plotting in plot.gam modified.

* plot.gam could fail with residuals=TRUE due to incorrect counting in 
  the code allowing use of termplot. plot.gam failed to prompt before a 
  newpage if there was only one smooth. gam and gamm .Rd files updated 
  slightly. 

1.1-3

* extract.lme.cov2 could fail for random effect group sizes of 1 
  because submatrices with only a row or column lose their dimensions, and 
  because single number calls to diag() result in an identity matrix. 

1.1-2

* Some model formulae constructed in interpret.gam and used in 
  facilitating safe prediction for parametric terms had the wrong 
  environment - this could cause gam to fail to find data when e.g. lm, 
  would find it. (thanks Thomas Maiwald)
  
* Some items were missing from the NAMESPACE file. (thanks Kurt 
  Hornik)
    
* A very simple formula.gam function added, purely to facilitate 
  better printing of anova method results under R 2.0.0. 

1.1-1

* Due, no doubt, to gross moral turpitude on the part of the author, 
  gamm() calculated the complete estimated covariance matrix of the 
  response data explicitly, despite the fact that this matrix is usually rather 
  sparse. For large datasets this could easily require more memory than 
  was available, and huge computational expense to find the choleski 
  decomposition of the matrix. This has now been rectified: when the 
  covariance matrix has diagonal or block diagonal structure, then this is 
  exploited.
    
* Better examples have been added to gamm().
    
* Some documentation bugs were fixed. 

1.1-0

Main changes are as follows. Note that `gam' object has been modified, so 
old objects will not always work with version 1.1 functions.

** Two new smooth classes "cs" and "ts": these are like "cr" and "tp" 
  but can be penalized all the way down to zero degrees of freedom to 
  allow fully automatic model selection (more self consistent than having a 
  step.gam function).
 
* The gam object expanded to allow inheritance from type lm and type 
  glm, although QR related components of glm and lm are not available 
  because of the difference in fitting method between glm/lm and gam.

** An anova method for gam objects has been added, for *approximate* 
  hypothesis testing with GAMs.
  
** logLik.gam added (logLik.glm with df's fixed): enables AIC() to be 
  used with gam objects.
  
** plot.gam modified to allow plotting of order 1 parametric terms via 
  call to termplot.
    
* Thanks to Henric Nilsson option `shade' added to plot.gam
    
* predict.gam modified to allow safe prediction of parametric model 
  components (such as poly() terms).
    
* predict.gam type="terms" now works like predict.glm for parametric 
  components. (also some enhancements to facilitate calling from 
  termplot())
    
* Range of smoothing parameter estimation iteration methods expanded 
  to help with non-convergent cases --- see ?gam.convergence
    
* monotonic smoothing examples modified in light of above changes.
    
* gamm modified to allow offset terms.
    
* gamm bug fixed whereby terms in a model formula could get lost if 
  there were too many of them.
    
* gamm object modified in light of changes to gam object. 

1.0-7

* Allows a model frame to be passed as `newdata' to predict.gam: it 
  must contain all the terms in the gam objects model frame, `model'.
    
* vis.gam() now passes a model frame to predict.gam and should be more 
  robust as a result. `view' and `cond' must contain names from 
  `names(x$model)' where x is the gam object. 

1.0-6/5/4

* partial residuals modified to be IRLS residuals, weighted by IRLS 
  weights. This is a much better reflecton of the influence of residuals 
  than the raw IRLS residuals used before.
    
* gamm summary sorted out by using NextMethod to get around fact that 
  summary.pdMat can't be called directly (not in nlme namespace exports).
    
* niterPQL and verbosePQL arguments added to gamm to allow more 
  control of PQL iteration.
    
* backquote=TRUE added when deparsing to allow non-standard names. 
  (thanks: Brian Ripley)
    
* bug in gam corrected: now gives correct null deviance when an offset 
  is present. (thanks: Louise Burt)
    
* bug in smooth.construct.tp.smooth.spec corrected: k=2 caused a 
  segfault as the C code was reseting k to 3 (actually null space 
  dimension +1), and not enough space was being allocated in R to handle 
  the resultng returned objects. k reset in R code, with warning. (Thanks: 
  Jari Oksanen)
    
* predict.gam() now has "standard" data searching using a model frame 
  based on a fake formula produced from full.formula in the fitted object. 
  However it also warns if newdata is present but incomplete. This means 
  that if newdata does not meet White book specifications, you get a 
  warning, but the function behaves like predict.lm etc. predict.gam had 
  been segfaulting if variables were missing from newdata (Thanks: Andy 
  Liaw and BR)
    
* contour option added to vis.gam
    
* te smooths can be forced to use only a single penalty (theoretical 
  interest only - not recommended for practical use) 

1.0-3

* Fixes bugs in handling graphics parameters in plot.gam()
    
* Adds option of partial residuals to plot.gam() 

1.0-2/1

* Fixes a bug in evaluating variables of smooths, knots and by-variables.

1.0-0

*** Tensor product smooths - any bases available via s() terms in a gam 
  formula can be used as the basis for tensor product smooths of multiple 
  covariates. A separate wiggliness penalty and smoothing parameter is 
  associated with each `marginal' basis.
    
** Cyclic smoothers: penalized cubic regression splines which have the 
  same value and first two derivatives at their first and last knots.
    
*** An object oriented approach to handling smooth terms which allows 
  the user to add their own smooths. Smooth terms are constructed using 
  smooth.construct method functions, while predictions from individual 
  smooth terms are handled by predict.matrix method functions.
    
** p-splines implemented as the illustrative example for the above in 
  the help files.
    
*** A generalized additive mixed model function gamm() with estimation 
  via lme() in the normal-identity case and glmmPQL() otherwise. The main 
  aim of the function is to allow a defensible way of modelling correlated 
  error structures while using a GAM.
    
* The gam object itself has changed to facilitate the above. Most 
  information pertaining to smooth terms is now stored in a list of smooth 
  objects, whose classes depend on the bases used. The objects are not 
  back compatible, and neither are the new method functions. This has been done 
  in an attempt to minimize the scope for bugs, given the amount of time 
  available for maintenance.
    
** s() no longer supports old stlye (version <0.6) specification of 
  smooths (e.g. s(x,10|f)). This is in order to reduce the scope for 
  problems with user defined smooth classes.
    
* The mgcv() function now has an argument list more similar to magic().
    
* Function GAMsetup() has been removed.
    
* I've made a general attempt to make the R code a bit less like a 
  simultaneous translation from C. 

0.9-5/4/3/2/1 

* Mixtures of fixed degree of freedom and estimated degree of freedom 
  smooths did not work correctly with the perf.iter=FALSE option. Fixed.
    
* fx=TRUE not handled correctly by fit.method="magic": fixed.
    
* some fixes to GAMsetup and gam documentation.
    
* call re-instated to the fitted gam object to allow updating
    
* -Wall and -pedantic removed from Makevars as they are gcc specific.
    
* isolated call to Stop() replaced by call to stop()! 

0.9-0 

*** There is a new underlying smoothing parameter selection method,
  based on pivoted QR decomposition and SVD methods implemented in LAPACK. 
  The method is more stable than the Wood (2000) method and allows the 
  user to fix some smoothing parameters while estimating others, 
  regularize the GAM fit in non-convergent cases and put lower bounds on 
  smoothing parameters. The new method can deal with rank deficient 
  problems, for example if there is a lack of identifiability between the 
  parametric and smooth parts of the model. See ?magic for fuller details. 
  The old method is still available, but gam() defaults to the new method.

* Note that the new method calls LAPACK routines directly, which means 
  that the package now depends on external linear algebra libraries,
  rather than relying entirely on my linear algebra routines. This is a 
  good thing in terms of numerical robustness and speed, but does mean 
  that to install the package from source you need a BLAS library installed 
  and accesible to the linker. If you sucessfully installed R by building 
  from source then you should have no problem: you have everything already 
  installed, but occasionally users may have to install ATLAS in order to 
  install from source.
    
* Negative binomial GAMs now use the families supplied by the MASS library 
  and employ a fast integrated GCV based method for estiamting the 
  negative binomial parameter. See ?gam.neg.bin for details. The new 
  method seems to converge slightly more often than the old method, and 
  does so more quickly.

* persp.gam() has been replaced by a new routine vis.gam() which is 
  prettier, simpler and deals better with factor covariates and at all 
  with `by' variables.
    
* NA's can now be handled properly in a manner consistent with lm() 
  and glm() [thanks to Brian Ripley for pointing me in the right direction 
  here] and there is some internal tidying of GAM so that it's behavious 
  is more similar to glm() and lm().
    
* Users can now choose to `polish' gam model fits by adding an nlm()  
  based optimization after the usual Gu (2002) style `power iteration' to
  find smoothing parameters. This second stage will typically result in a 
  slightly lower final GCV/UBRE score than the defualt method, but is much 
  slower. See ?gam.control for more information.
    
* The option to add a ridge penalty to the GAM fitting objective has been 
  added to help deal with some convergence issues that occur when the
  linear predictor is essentially un-identifiable. see ?gam.control. 

0.8-7

* There was a bug in the calculation of identifiability side conditions 
  that could lead to over constraint of smooths using `by' variables in
  models with mixtures of smooths of different numbers of variables. This 
  has been fixed. 

0.8-6

* Fixes a bug which occured with user supplied smoothing parameters, in 
  which the weight vector was omitted from part of the influence (hat) 
  matrix calculation. This could result in non-sensical variance 
  estimates.
    
* Stronger consistency checks introduced on estimated degrees of freedom.

0.8-5

* mgcv was using Machine() which is deprecated from R 1.6.0, this 
  version uses .Machine instead. 

0.8-4 

* There was a memory bug which could occur with the "cr" basis, in 
  which un-allocated memory was written to in the tps_g() routine in the 
  compiled C code - this occured when that routine was asked to clean up 
  its memory, when there was nothing to clean up. Thanks to Luke Tierney for 
  finding this problem and locating it to tps_g()!
    
* A very minor memory leak which occured when knots are used to start 
  a tps basis was fixed. 

0.8-3 

* Elements on leading diagonal of Hat/Influence matrix are now 
  returned in gam object.
    
* Over-zealous error trap introduced at 0.8-2, caused failure with 
  smoothless models. 

0.8-2

* User can now supply smoothing parameters for all smooth terms (can't 
  have a mixture of supplied and estimated smoothing parameters). Feature 
  is useful if e.g. GCV/UBRE fails to produce sensible estimates.
    
* svd() replaced by La.svd() in summary.gam(). 
    
* a bug in the Lanczos iteration code meant that smooths behaved 
  poorly if the smooth had exactly one less degree of freedom than the 
  number of data (the wrong eigenvectors were retained in this case) - 
  this was a rather rare bug in practice!
    
* pcls() was not using sensible tolerances and svdroot() was using 
  tolerances incorrectly, leading to problems with pcls(), now fixed.
    
* prior weights were missing from the pearson residuals.

* Faulty by variable documentation fixed (have lost name of person who 
  let me know this, but thanks!)
    
* Scale factor removed from Pearson residual calculation for 
  consistancy with a higher proportion of authors.
    
* The proportion deviance explained has been added to summary.gam() as 
  a better measure than r-squared in most cases.
    
* Routine SANtest() has been removed (obsolete).
    
* A bug in the select option of plot.gam has been fixed. 

0.8-1 

* The GCV/UBRE score can develop phantom minima for some models: these 
  are minima in the score for the IRLS problem which suggest large 
  parameter changes, but which disappear if those large changes are 
  actually made. This problem occurs in some logistic regression models. 
  To aid convergence in such cases, gam.fit now switches to a cautious 
  mgcv optimization method if convergence has not been obtained in a user 
  defined number of iterations. The cautious mode selects the local 
  minimum of the GCV/UBRE closest to the previous minimum if multiple 
  minima are present. See gam.control for details about controlling 
  iterations.
    
* Option trace in gam.control now prints and plots more useful 
  information for diagnosing convergence problems.
    
* The one explicit formation of an inverse in the underlying multiple 
  GCV optimization has been replaced with something more stable (and 
  quicker).
    
* A bug in the calculation of side conditions has been fixed - this 
  caused a failure with models having parametric terms and terms like: 
  s(x)+s(z)+s(z,x).
    
* A bug whereby predict.gam simply failed to pick up offset terms has 
  been fixed.
    
* gam() now drops unused levels in factors.
    
* A bug in the conversion of svd convergence criteria between version 
0.7-2 and 0.8-0 has been fixed.

* Memory leaks have been removed from the C code (thanks to the superb 
  dmalloc library).
    
* A bug that caused an undignified exit when 1-d smoothing with full 
  splines in 0.8-0 has been fixed.

0.8-0 

* There was a problem on some platforms resulting from the default 
  compiler optimizations used by R. Specifically: floating point registers  
  can be used to store local variables. If the register is larger than a 
  double (as is the case for Intel 486 and up), this means that:
      double a,b;
      a=b;
      if (a==b)
  can evaluate as FALSE. The mgcv source code assumed that this could 
  never happen (it wouldn't under strict ieee fp compliance, for example). 
  As a result, for some models using the package compiled using some 
  compiler versions, the one dimensional "overall" smoothing parameter 
  search could fail, resulting in convergence failure, or undersmoothing. 
  The Windows version from CRAN was OK, but versions installed under Linux 
  could have problems. Version 0.8 does not make the problematic 
  assumption.
    
* The search for the optimal overall smoothing parameter has been 
  improved, providing better protection against local minima in the 
  GCV/UBRE score.
    
* Extra GCV/UBRE diagnostics are provided, along with a function 
  gam.check() for checking them.
    
* It is now possible for the user to supply "knots" to be used when 
  producing the t.p.r.s. basis, or for the cubic regression spline basis. 
  This makes it feasible to work with very large datasets using the 
  of the data. It also provides a mechanism for obtaining purely "knot 
  based" thin plate regression splines.
    
* A new mechanism is provided for allowing a smooth term to be 
  multiplied by a covariate within the model. Such "by" variables allow 
  smooths to be conditional on factors, for example.
    
* Formulae such as y~s(x)+s(z)+s(x,z) can now be used.
    
* The package now reports the UBRE score of a fitted model if UBRE was 
  used for smoothing parameter selection, and the GCV score otherwise.
    
* A new help page gam.models has been added.
    
* A bug whereby offsets in model formulae only worked if they were at 
  the end of the formulae has been fixed.
    
* A bug whereby weights could not be supplied in the model data frame 
  has been fixed.
   
* gam.fit has been upgraded using the R 1.5.0 version of glm.fit
    
* An error in the documentaion of xp in the gam object has been fixed, 
  in addition to numerous other changes to the documentation.
    
* The scoping rules employed by gam() have been brought into line with 
  lm() and glm by searching for variables in the environment of the model 
  formula rather than in the environment from which gam() was called - 
  usually these are the same, but not always.
    
* A bug in persp.gam() has been fixed, whereby slice information had 
  to be supplied in a particular order.
    
* All compiled code calls now specify package mgcv to avoid any 
  possibility of calling the wrong function.
    
* All examples now set the random number generator seed to facilitate 
  cross platform comparisons. 

0.7-2

* T and F changed to TRUE and FALSE in code and examples.
    
* Minor predict.gam error fixed (didn't get correct fitted values if 
  called without new data and model contained multi-dimensional smooths). 

0.7-1

* There was a somewhat over-zealous warning message in the single 
  smoothing parameter selection code - gave a warning everytime that GCV 
  suggested a smoothing parameter at the boundary of the search interval - 
  even if this GCV function was also flat. Fixed.
    
* The search range for 1-d smoothing parameter selection was too wide 
  - it was possible to give so little weight to the data that numerical 
  problems caused all parameters to be estimates as zero (along with the 
  edf for the term!). The range has been narrowed to something more sensible 
  [above warning should still be triggered if it is ever too narrow - but 
  this should not be possible].
    
* summary.gam() documentation extended a bit. p-values for smooths are 
  slightly improved, and an example included that shows the user how to 
  check them! 

0.7-0

* The underlying multiple GCV/UBRE optimization method has been 
  considereably strengthened, as follows:
  o First and second guess starting values for the relative 
    smoothing parameters have been improved.
  o Steepest descent is used if either: i) the Hessian of the 
    objective is not positive definite, or (ii) Steps in the Newton direction 
    fails to improve the GCV/UBRE score after 4 step halvings (since in 
    this case the quadratic model is clearly poor).
  o Newton steps are rescaled so that the largest step component 
    (in log relative smoothing parameters) is of size 5 if any step 
    components are >5. This avoids very large Newton steps that can occur 
    in flat regions of the objective.
  o All steepest descent steps are initially scaled so that their 
    longest component is 1, this avoids long steps into flat regions of 
    the objective.
  o MGCV Convergence diagnostics are returned from routines mgcv 
    and gam.
  o In gam.fit() smoothing parameters are re-auto-initialized 
     during IRLS if they have become so far apart that some are likely to 
     be in flat parts of the GCV/UBRE score.
  o A bug whereby poor second guesses at relative smoothing 
    parameters could lead to acceptance of the first guess at these 
    parameters has been removed.
  o The user is warned if the initial smoothing parameter guesses 
    are not improved upon (can happen legitmately if all s.p.s should be 
    very high or very low.) 
      
  The end result of these changes is to make fits from gam much more 
  reliable (particularly when using the tprs basis available from version 
  0.6).

* A summary.gam and associated print function are provided. These 
  provide approximate p-values for all model terms.
    
* plot.gam now provides a mechanism for selecting single plots, and 
  allows jittering of rug plots.
    
* A bug that prevented models with no smooth terms from being fitted 
  has been removed.
    
* A scoping bug in gam.setup has been fixed.
    
* A bug preventing certain mixtures of the bases to be used has been 
  fixed.
    
* The neg.bin family has been renamed neg.binom to avoid masking a 
  function in the MASS library. 

0.6-2 
revisions from 0.6.1

* Relatively important fix in low level numerics. Under some circumstances 
  the Lanczos routines used to find the thin plate regression spline basis 
  could fail to converge or give wrong answers (many thanks to Charles 
  Paxton for spotting this). The problem was with an insufficiently stable 
  inverse iteration scheme used to find eigenvectors as part of the 
  Lanczos scheme. The scheme had been used because it was very fast: 
  unfortuantely stabilizing it is as computationally costly as simply 
  accumulating eigen-vectors with the eigen-values - hence the latter has 
  now been done. Some further examples also added. 

0.6-1

* Junk files removed from src directory. 

* 3 C++ style comments removed from tprs.c.

0.6-0

* Multi-dimesional smoothing is now available, using "thin plate 
  regression splines" (MS submitted). These are based on optimal 
  approximations to the thin-plate splines.
    
* gam formula syntax upgraded (see ?s ). Old syntax still works, with 
  the exception that if no df specified then the tprs basis is always used 
  by default.
    
* plot.gam can now deal with two dimensional smooth terms as well as 
  one dimensional smooths.
    
* persp.gam added to allow user to visualize slices through a gam 
  [Mike Lonergan]
    
* negative binomial family added [Mike Lonergan] - not quite as robust 
  as rest of families though [can have convergence problems].
    
* predict.gam now has an option to return the matrix mapping the 
  parameters to the linear predictor at the supplied covariate values.
    
* Variance calculation has been made more robust.
    
* Routine pcls added, for penalized, linearly constrained optimization 
(e.g. monotonic splines).
    
* Residual method provided (there was a bug in the default - Thanks 
  Carmen Fernandez).
    
* The cubic regression spline basis behaved wrongly when extrapolating 
  [thanks Sharon Hedley]. This is now fixed.
    
* Tests included to check that there are enough unique covariate 
  combinations to support the users choise of smoothing basis dimension.
    
* Internal storage improved so that large numbers of zeroes are no 
  longer stored in arrays of matrices.
    
* Some method argument lists brought into line with the R default 
  versions. 

0.5

    
* There was a bug in gam.fit(). The square roots of the correct iterative 
  weights were being used in place of the weights: the bug was
  apparent because the sum of fitted values didn't always equal the sum of 
  the response data when using the canonical link (which it should as a 
  result of X'f=X'y when canonical link used and unpenalized). The bug has 
  been corrected, and the correction tested. This problem did not affect 
  (unweighted) additive models, only generalized additive models.
    
* There was a bug that caused a crash in the compiled code when there were 
  more than 8000 datapoints to fit. This has been fixed.
    
* The package now reports its version number when loaded into R.
    
* predict.gam() now returns predictions for the original covariate values 
  (used to fit the model) when called without new data.
    
* predict.gam() now allows type="response" as an argument - returning 
  predictions on the scale of the response variable.
    
* plot.gam() no-longer defaults to automatic page layout, use argument 
  pages=1 to get the old default behaviour.
    
* A bug that could cause a crash with the model formula y~s(x)-1 has been 
  fixed.
    
* Yet more sloppy practices are now allowed for naming variables in model 
  formulae. e.g. d$y ~ s(d$x) now works, although its not recommended.
    
* The GCV score is now reported by print.gam() (whether or not GCV was 
  actually used - it isn't the default for Poisson or binomial).
    
* plot.gam() modified to avoid prompting for input when not used 
  interactively.

0.4 

* Transformations allowed on lhs of gam formulae .
    
* Argument order same as Splus gam.
    
* Search for data now designed to be like lm() , so you can now be quite 
  sloppy about where your data are.
    
* The above mean that Venables and Ripley examples can be run without 
  having to read the documentation for gam() so carefully!
    
* A bug in the standard error calculations for parametric terms in 
  predict.gam() is fixed.
    
* A serious bug in the handling of factors was fixed - it was previously 
  possible to obtain a rank deficient design matrix when using factors, 
  despite having specified an identifiable model.
    
* Some glitches when dealing with formulae containing offset() and/or I() 
  have been fixed.
    
* Fitting defaults can now be altered using gam.control when calling gam()

0.3-3
    
* Documentation updated, including removal of wrong information about 
  constraints and mgcv . Also some readability changes in code and no 
  smooths are now allowed.
    
0.3-2/1

* Allows all ways of specifying a family that glm() allows (previously 
  family=poisson or family="poisson" would fail). Some more documentation 
  fixes.
    
* 0.2 lost the end of long formulae (because of a difference in the way 
  that R and Splus deal with formulae). This is now fixed.
    
* A minor error that meant that QT() failed under some versions of Windows 
  is now fixed.
    
* All package functions now have help(). Also the help files have been 
  more carefully checked - version 0.2 actually contained no information 
  on how to write a GAM formula as a result of a single missing '}' in the 
  help file!

0.2

* Fixed d.f. regression splines allowed as part of gam() model 
  specification.
    
* Bug in knot placement algorithm fixed (caused crash with df close to 
  number of data).
    
* Replicate covariate values dealt with properly in gam()!
    
* Data search method in gam() revised - now looks in frame from which 
  gam() called.
    
* plot.gam() can now deal with missing variance estimates gracefully.
    
* Low (1,2) d.f. smooths dealt with gracefully by gam() - no longer cause 
  freeze or crash.
    
* Confidence intervals simulation tested for normal(identity), 
  poisson(log), binomial(logit) and gamma(log) cases. Average coverage 
  probabilities from 0.89 to 0.97 term by term, 0.93 to 0.96 "across the 
  model", for nominal 0.95.
    
* R documentation updated and tidied.