Ox 4.0: new features and fixes

New features and enhancements
Fixed problems and new features in 4.1
Problems fixed in 4.04
Problems fixed in 4.02
Problems in version 3.40 fixed in 4.00
OxGauss

New features and enhancements

Added switch and switch_single language constructs (and keywords).
Direct access to data member variables of a class can now be provided (struct, public, protected). In a struct all members are public. In a class, public can be used to make data public (but protected and public do not apply to functions).
New default uniform random number generator (MWC8222_52) with period 2^8222 and using full 52-bit floating point precision. Also new (and faster) standard normal generator based on the ziggurat method. Use ranseed("PM") to reset the old default for uniform and normal, in order to replicate results from older Ox versions.
New FFT function fft1d, which is faster and works for arbitrary sample size. The original fft now uses the faster FFT version, and still converts the sample size to a power of two. Modified dfft to make it compatible with fft1d/fft1d. The new FFT is based on FFTE 2.0 by Daisuke Takahashi.
dayofeaster, dayofmonth, timeofday, "%C" format.
W test for normality (NormTestW in lib/normtest).
Added polyeval.
MaxSQP: removed singularity message in Maratos correction (now QR based), and many other improvements.
%t and %T sscan/fscan formats to scan a token
%v sscan/fscan and print/fprint formats to read and write values a Ox constant (literals). This can be used, e.g. to read and write the entire contants of a class object. An example is in samples/inout/percent_v.ox.
Small improvement to format handling in print/fprint makes it easier to print strings with % in it. For example in print("%s", "%GDP") the second argument is treated as a string to print, not as a format.
Database class uses findsample in SetSelSample; added GetSelSampleMode/SetSelSampleMode; added Save and SetVar. Added support for daily data (dated database). Ability to read Excel files with date/time columns.
Extensions to Modelbase class to support many enhancements of OxPack.
Added findsample; loadmat/savemat can handle variable names
Slightly more accurate tailn/probn for probablities > 1e-20 (about 14-15 significant digits correct, instead of 10-11 before).
Ox can run Ox files that use non-English characters in the filename (Windows 2000/XP).
The default symbol table size is now 3000.

Fixed problems and new features in version 4.1:

Improved PcFiml output for single equation models.
10^10 as a constant expression (i.e. outside any function) would overflow (using integer arithmetic). Now works same as the run-time expression.
Modelbase::DoOptionsDlg calls SaveOptions() when successful.
The 64-bit version of Ox defines OX_64_BIT. Also, the 64-bit version will first try xyz_64 (with the appropriate .so or .dll extension) if the dynamic link library is specified as xyz in the extern statement. So now, e.g., arfima.so and arfima_64.so can be installed in the arfima folder, and the correct version will be used automatically.
Fixed several problems with loading CSV files.
densbinomial, densgeometric, densnegbin returned .NaN with p=1 or p=0.
Random number generators for distributions with two parameters (such as rangamma) would in some circumstances use the second argument as the third. This would only happen if they are both matrices, conformant to the dimension of the generated matrix (so not if both are scalars, the most common case).
The following code: gave a runtime error: 'matrix[1][2]' cannot cast to int while it should just work.
Select Variables dialog (OP_SELECTVAR) would allow database changes based on the previous incantation rather than the current. The actual database change also didn't work.
GetSampleByIndex in help: link does not work
countc could crash; now the second argument in countc/countr can be a column as well as a row vector
Fixed some corner cases of probf/tailf: argument 0 always returns 0 (1 for tailf), +.Inf returns 1 (0 for tailf).
quanbeta breaks down for large values of a and b (when distribution becomes quite rectangular). This was partially because of bad starting values.
unvech with a scalar argument returns 0 instead of a 1x1 matrix with the scalar.
reading xls file with embedded chart data (perhaps without actually holding a chart) could go wrong.
Missing values for min/max in DrawAxis causes infinite loop
PcFiml class: Chow test didn't work in a model without regressors.

Problems which have been fixed in version 4.04:

Ox Console install in OxEdit: added Ox - compile command to Run menu.
Changed .oxo link error to a warning if a version 3 compiled oxo file is used under version 4. Note that this will not help with code that imports a class that has changed (Database, Modelbase e.g.).
countc/countr now skip over missing values.
savemat for .mat/.fmt file: now ignores third argument if it is an array (previously generated a run-time error).
Loading Excel files:
Problem: if the first column looks like it defines a fixed frequency sample and has no label, it is not loaded. This has been fixed: the first column is maintained if it has a label or if it has numeric data. (same as GiveWin and Ox 3, Ox 4.00/4.02 would drop it).
Enhancements:
- A dot or any text starting with #N/A is treated as a missing value (previously it was only #N/A exactly).
- Now dropping columns which have no entries at all (as happened in Ox 3/GiveWin)
- loadmat defaults to treating strings as missing values (this is as in Ox 3, but different from loading an Excel file using the Database class). Dates are also left untranslated by default. This maintains symmetry between loadmat and savemat.
- If the format argument of loadmat is set to 1 for an xls file, strings and dates are read in the same way as using the Database class. So loadmat("x.xls") is now the same as Ox 3.x), while loadmat("x.xls", 1) maintains strings as numbers and translates dates).
deletec(x,<>) and selectc(x,<>) crash on empty second argument.
insertc/insertr at the position of one plus the last column/row crashes.

Problems in version 4.00 which have been fixed in version 4.02:

OxPack:
- main help command was not working
- No OP_ESTIMATE or OP_ESTIMATE_NODB: always switch test menu entries on.
Modelbase::DbDrawTMatrix did not grow the database correctly when required.
eigensym: possible error in eigenvectors of blockdiagonal symmetric matrix.
eigensymgen: now much faster for large matrices.
switch/switch_single which doesn't end with default leads to IE Stack Error if the last case is executed and FALSE. CTL_SAMPLEINDEX field inside OP_MATRIX_CTL: entering a number before the inline edit field is shown leads to crash (is fine in all other control types)
getenv will (most likely) lead to a crash
64-bit linux only: using a log file can lead to a crash when printing strings with string format specifier.
g2ox.ox: loaddata and savedata definition wrong

Problems in version 3.40 which have been fixed in version 4.00:

Fixed problem with: delete x[i] where x is an array of objects.
Fixed: accessing a baseclass static variable from inside a derived class did not work, because it behaved as if the derived class had its own copy. Using the baseclass's access function (if written) would work correctly though.
this->undeclared[someindex] = somevalue; would lead to crash in parsing after several warning messages
Fixed problem with destructor within namespace
trace of a scalar double is wrong (so no problem with trace(<1>) but with trace(1.0)). cdiv(< .;1 >,< .;1 >) would give divide by zero warning. Now it returns < .;. >.
The following should work but gave an error message that a scalar was expected:
countr, countc: if the intervals are not ordered, the return value is filled with missing values.
sortcindex(<>) crashes.
Cannot load an in7/bn7 data file which has only one observation.
MaxSQP could (rarely: when vc.==0) get stuck in an infinite loop in the computation of the slope w.r.t. constraints in merit function. Also improved MaxSQP line search.
SolveQP_chol could (very rarely) get stuck in an infinite loop when the ddiag value is zero.

OxGauss

Added support for varput/varget.
Fixed a bug has been identified in a procedure which (1) calls as yet undefined procedures, (2) includes two or more for loops. In that case, the second (and further) loop incrementation can go awry. The reason is that there is insufficient space for the loop-local variables on the stack.
Fixed bug in y/x when it means regression and trap(2) is on (replacing missing values with 0). ftos with character matrix didn't work:

This file last changed .