ChangeLog

Since April 2014, this changelog includes only those changes that break existing code or are especially notable. For a full list of changes, see the Git history at 
https://github.com/b-k/Apophenia/commits/master


[Key, pre-March 2014: 
--Addition or improvement
**Change that could require recoding existing code.
!!Big.
]

    October 2014
** apop_model_stack --> apop_model_cross

    August 2014
** default for apop_data_pack is .all_pages='y' (was 'n').
** remove apop_plot_lattice, apop_plot_triangle, apop_plot_line_and_scatter, apop_plot_qq.
Find them at https://github.com/b-k/apophenia/wiki/gnuplot_snippets .

	March 2014
--Command-line tools print help should a user add a --help option.

	February 2014
!!OpenMP for threading. All calls to apop_map and friends, apop_model_draws, and others auto-thread.
!!apop_rng_get_thread to get a thread-specific RNG, so you can thread random processes.

	January 2014
**View macro reform:
	Apop_cols       Apop_rows       A contiguous set of columns or rows as an apop_data set (with names)
	Apop_col	    Apop_row        One column or row as an apop_data set
	Apop_col_t      Apop_row_t      One column or row as an apop_data set, retrieved by row/col name
	Apop_col_v      Apop_row_v      One column or row as a gsl_vector
	Apop_col_tv     Apop_row_tv     One column or row as a gsl_vector, retrieved by row/col name
	Apop_matrix_col Apop_matrix_row One column or row as a gsl_matrix
**MLE methods are now strings instead of all-caps enums.
**All blank elements of a data->text grid point to the same NUL string.
--Add apop_model_metropolis; revise apop_update accordingly.
--apop_draw uses metropolis to draw from any model with a log likelihood/p and where data size>1.
**Replace all instances of output_file with output_name (GNU sed -i 's/output_file/output_name/g' *c)
--Consolidated headers
**apop.h no longer #includes time.h or unistd.h
**apop_draw returns zero on success; nonzero on failure.
**Removed the BIC-by-cells from estimation output. Added AIC_c. OLS now reports the ICs along with R^2.

	December 2013
--append and replace options for apop_text_to_db
--apop_probit bug fix
**apop_plot_histograms now uses gnuplot's impulses, not boxes by default---handles missing zero bins better.
**MLE path trace lists the probs/loglikelihoods in the vector of the apop_data set it produces. path is apop_data**, from apop_data*.
**apop_data_transpose has an .inplace option, which is 'y' by default. Add .inplace='n' to existing uses.
--siman checks constraints for the starting point.
--Mixture models overhauled.
--cleaned up the command-line utilities
**removed apop_lookup.

	November 2013
--rewrite apop_data_sort to allow sorting by multiple columns or text or names
--apop_pmf now has a CDF method.
--fixed up K-S tests.
--removed the Swig Python wrapper from this package.
**replaced char apop_opts.db_nan[101] with char *apop_opts.nan_string. More descriptive, easier to use.
**apop_name_find does plain case-insensitive search; no regexes.

	October 2013
**all built-in models (apop_ols, apop_dirichlet, ...) are now apop_model* (ptr-to-struct), from apop_model (plain struct).
**apop_estimate and apop_copy take in an apop_model* instead of plain apop_model.
**printing no longer part of the apop_model struct; uses a vtable.

	September 2013
**change vbase, m1base, m2base ==> vsize, msize1, msize2
**Estimate returns void (was apop_model*)
--vtable mechanism improvements
**Remove score, predict, and parameter_model from the apop_model object; use the vtable mechanism.
**Upgrade model p, ll, cdf, constraint to return long double (was double)
**consolidate vector_var and vector_weighted_var. same with cov, mean, weighted_skew, and weighted_pop. Users just have to replace apop_vector_weighted_var with apop_vector_var.
**removed deprecated.h entirely.
--apop_data_add_named_elmts puts new data in the vector, not the matrix, because it is intended for a list of scalars (==a vector). If you use apop_data_get(infodata, .rowname="statistic name") then you'll be able to retrieve the element either way.
**removed apop_line_to_data and apop_line_to_matrix. Use apop_data_fill and apop_data_falloc.

	August 2013
--apop_map(_sum) properly threads data-row mappings. .inplace='v' to return NULL.
**Remove apop_settings_alloc, apop_settings_group_alloc
**Change Apop_row to return an apop_data set, not a vector (for which use Apop_matrix_row).
**Apop_settings_set sets model->error='s' on error, instead of returning.
**Add a .want_path='y' setting to your apop_mle_settings group, and I'll put a list of the points tried by the optimizer in an apop_data set named path (in the settings group); see documentation for details. Remove the former trace_path mechanism.
**removed apop_(vector|matrix)_increment. Use, e.g., *gsl_vector_ptr(v, 7) += 3; or (*gsl_vector_ptr(v, 7))++.
**Some mu and sigmas => μ and σ
**Removed apop_settings_alloc, apop_settings_group_alloc
**Change Apop_row to return an apop_data set, not a vector (for which use Apop_matrix_row).

	June 2013
--replaced makefile in base directory with ./configure.
--version number now equals build date.
**name->title is a ptr; name->column => name->col
**removed apop_strip_dots; it's up to the user to give reasonable names for the db.

	May 2013
--jacobian transformations
--Apop_model_copy_set to copy a model and add a settings group at once
--mixture models
--data-data composition
--added apop_model_draws
!!vtables, allowing for more functions with special cases for certain model(s) outside of the model object itself.

	April 2013
--plugged some memory leaks
--default tolerance for MLE is much finer (1e-5).
--Added apop_text_fill
**Finally removed support for gsl_histograms, including the apop_histogram model. This cut has been promised for about four years now. Use the apop_pmf instead.
**apop_data_to_bins no longer modifies the input data in place. It now makes a copy and modifies the copy.
**Removed apop_crosstab_to_pmf. There's a version at https://github.com/b-k/Apophenia/wiki/Crosstab-to-PMF for matrices.
**Removed apop_vector_to_array. If your array has stride 1, use your_array->data; else write a for loop to copy out the data.
**Removed apop_array_to_matrix.

	March 2013
**apop_text_paste now prints the pasted string at verbosity level 3 (formerly 2).

	February 2013
--Starting_point in Bayesian updating no longer does anything, but it was never significant to begin with. Added some verbosity options to apop_update.

	January 2013
--Logit regression much smarter about picking a starting point.
--Defaults for simulated annealing try 1600x fewer points. Prior settings were overkill.
--configure.ac checks for native asprintf and uses it if it is present
**Removed apop_db_merge and apop_db_merge_table. Get them from http://modelingwithdata.org/arch/00000141.html .
**Removed apop_matrix_fill; use apop_data_fill
**Removed apop_array_to_data.
**Removed apop_matrix_correlation; use apop_data_correlation
**Deprecated apop_rank_compress; use apop_data_to_dummies(..., .keep_first='y').
--apop_model_stack

	December 2012
--Added an apop_pmf_settings group, eliminating a couple of hacks (e.g., see October 2012).
--faster read-in of text files
--Exponential model uses data in both the vector and matrix parts of the apop_data input
--transformation to generate mixtures (i.e., linear combinations) of models.
**apop_data_transpose now transposes the text element as well as the matrix (by default).
	Use apop_data_transpose(your_matrix,  .transpose_text='n') to replicate the previous behavior.
--removed Autoconf pkg-config macros, because Autoconf no longer needs the help.
--writing apop_data to DB uses prepared queries where possible => much faster.
**It is now up to you to put apop_query("begin"); / apop_query("commit"); wrappers around
	writing of tables to the database.

	November 2012
--apop_vector_unique_elements, apop_data_to_dummies, and apop_data_to_factors handle NaNs
	better; put them at the end of the sort order.
!!Finally added an .error element to apop_data and apop_model structs, thus simplifying
	error-checking.
--Apop_stopif macro, rendering the Apop_assert family largely obsolete (so if you're using
	them in your own work, consider them deprecated...).
--Where the assert macros used to abort() on errors, they now send signal(SIGTRAP), making
	debugging a little easier. Most host systems force an abort on SIGTRAP anyway.

	October 2012
**Removed apop_strcmp. If you still need it, this macro is basically equivalent:
    #define apop_strcmp(a, b) (((a)&&(b) && !strcmp((a), (b))) || (!(a) && !(b)))
--clean up of parameter-fixing model transformation.
--split off multiple imputation variance code.
**Removed apop_vector_grid_distance. Use apop_vector_distance(v1, v2, .metric='M');
--If apop_pmf.dsize==0, apop_pmf.draw returns a row number, not the data in that row. This will change shortly.
--Logit draw function, akin to apop_ols.draw. Both will change shortly.
--apop_data_to_factors now auto-allocates a matrix if need be (because it always auto-allocated a vector). 

	September 2012
--\0 in text files counts as white space
--fixed counting bug for text files with ,<end-of-line> sequences.

	July 2012
--some MLE cleanup
--fixes to apop_rake
--apop_logit.score fixed

	June 2012
--The sample kurtosis calculation is still more precise.

	May 2012
--Autotools improvements. Use the standard 'make check' instead of the ad hoc 'make test'.
!!Set apop_opts.stop_on_warning='n' to never abort() on any type of error. E.g., GUIs that
	should never halt will use this. Default is still to halt on errors, because that's
	most useful for interactively developing numeric analyses.
--Use apop_opts.log_file=fopen("yourlog", "w") to divert the warnings/errors from stderr into yourlog.
--Some formerly void functions now return an int, to return an error code. E.g.,
	apop_opts.stop_on_warning='n';
	if (apop_data_set(data, row, col)) printf("Error! Nothing was set.\n");
--Fixed a memory leak in simulated annealing.
--Apop_data_row and apop_data_set_row handle row names

	March 2012
--bug fix in apop_text_to_data when input file has no names.
--probit dlog likelihood isn't implemented for N>2; now acknowledging this.
--added apop_data_get_factor_names
--apop_(vector|matrix)_(map|apply) now accepts NULL input.

	January 2012
**Removed apop_matrix_var_m, which nobody was using.
--bug fix in apop_vector_distance for Ln norms where n is odd.
--reading data from text files rewritten; much more robust. 
   **A space is no long a default delimiter.  Use apop_opts.input_delimiters="| ,\t" to restore old default.
   **apop_opts.db_nan is no longer a regex; I just do a case-insensitive comparison.

	December 2011
--apop_model.textsize is now a size_t instead of an int.
--apop_update accepts likelihoods with no pre-set parameters 

	November 2011
--moved to Github; some changes to structure and documentation to accommodate. 
**apop_ols.predict didn't do the OLS shuffle if the input has no vector; this was anomalous.

	October 2011
--apop_text_paste added.
**apop_multinomial and apop_binomial overhauled. No longer accepting Bernoulli draws as input.
--standardization: make docs => make doc

	September 2011
--`query turned up a blank table' warning turns up when apop_opts.verbose >=2. (used to be >=1)
--apop_t_test and apop_paired_t_test are quieter---no intermediate results until apop_opts.verbose >=2.
**apop_opts.db_name_column now has a blank default (instead of the SQL-specific and potentially surprise-inducing "rowname").

	August 2011
--Bootstrap/Jackknife are better with text
**apop_data_memcpy used to reallocate memory for the text and names elements; use apop_data_copy if you want allocation done for you.

	July 2011
--Apop_data_rm_rows now accepts a test function as well as a fixed list of rows to drop.

	June 2011
--apop_crosstab_to_db handles missing labels and NaNs better.
**apop_matrix_to_db removed (as promised a few years ago). Use apop_matrix_print(yourdata, "tabname", .output_type='d').
**F-test defaults now match ANOVA tradition.
--documentation script doesn't use GNU extensions to awk; should now be POSIX-standard.

	May 2011
**apop_map returns a data set with an allocated/filled vector when not called with .inplace='y'.  Before, it had been making a full copy, which is idiosyncratic.
--apop_rake accepts a weights column.
--apop_anova uses variadic arguments for a marginally nicer interface (and better argument checking).
**apop_data_to_dummies tries to give nicer labels. You may have to recode things if you relied on the old labels.

	March 2011
--Header files have been merged. A few long files is as easy to grep as a multitude of
	nearly one-line files. If you #include <apop.h> instead of the individual headers,
	then this shouldn't affect you. Due to redundancy, compilation with gcc takes 3% longer.

	0.99 February 2011
!!The apop_PMF model has more support:
	--New supporting functions: apop_data_pmf_compress, apop_model_to_pmf
	--functions that took apop_histogram models now take apop_pmfs as well:
		apop_test_chi_squared_goodness_of_fit, apop_test_kolmogorov
	--Consider the apop_histogram to be deprecated. Only two associated functions were removed; see below.
	**apop_histogram_plot is removed. Replace with:
	    fprintf(apop_opts.output_pipe, "plot '-' using 1:3 with boxes\n");
		apop_model_print(hist);
		fprintf(apop_opts.output_pipe, "e\n");
	**apop_histogram_print was a bad idea to begin with, because it basically replicates
	    gsl_histogram_fprintf. Use apop_model_print(your_histogram), which calls gsl_histogram_fprintf, 
		or call that function directly. The only difference: the GSL function prints 
		[start of bin] [end of bin] [value]
		and apop_histogram print showed
		[start of bin] [value]

	December 2010
**apop_maximum_likelihood no longer calls apop_prep. If you want that, use apop_estimate.
--apop_text_to_db lets users specify types and keys.
--deleted some obsolete/deprecated items: apop_error, apop_multinomial_settings
--apop_data_split retains text when splitting by rows; still ignores it when splitting by columns.

	November 2010
--apop_listwise_delete uses apop_opts.db_nan to check for missing data in the text part of the input data.
--apop_data_split handles names

    September 2010
**Multinomial distribution sets N to be the length of the row (a single observation)
    rather than the size of the full data set. Added apop_multinomial.parameter_model method
    for testing purposes.

    August 2010
** What was apop_assert => apop_assert_c; what was apop_assert_s => apop_assert. Their
    arguments are slightly different, and the thing that was asserted no longer prints along
    with the message you chose.
--verbosity defaults to 1. Queries print at verbosity >=2.
--apop_data_to_db writes the weights.
--Iterative proportional fitting, aka raking. 

**apop_text_add now frees the contents of cell in the text grid that you are about to
    overwrite, thus preventing memory leaks without effort from the user. If your existing
    code has other pointers to the string in that text cell, you'll have to replace the now
    string-freeing apop_text_add with asprintf(&(your_dataset->text[row][col]), "your string").

    July 2010
** apop_regex now gets all matches when you pull substrings. Each row of the text grid
    is a match, and if you have multiple substrings, each match's substrings will be
    along the columns.  May require recoding because the substrings used to be along
    the rows; just switch the indices.

    June 2010
** Removed the apop_rank settings group, and thus all the code related to it. It was just
    the wrong place to do this. Added apop_data_rank_expand to convert rank data to
    what the various models typically expect. This is another step for some users and
    could be a problem if the counts get into the billions, but it still makes more sense
    than rewriting every model twice.

    May 2010
**apop_data_prune_columns_base now takes in a list of strings terminated by a NULL, not by
a zero-length string.
--apop_data_get_row lets you pull a view from a data set. [this was briefly the apop_data_row]
    ==>apop_data_set_rows, apop_data_rm_rows
        ==> apop_data_listwise_delete is fifty lines shorter.
--apop_parts_wanted_settings: fixes some infinite loops (est needs parameter models ->
    p.m. bootstraps for variances -> bootstrap runs estimate -> repeat) and allows
    just-the-parameters estimation when you want it.
--cleaned up build system, including an added RPM spec file

    April 2010
--apop_t_distribution now has three parameters: mean, std dev, df. That is, it is based on un-normalized data.
**apop_random_int and apop_random_double removed for not being particularly useful.

    March 2010
**The apop_predict special case for when all data is non-missing was a bit too special,
    and has been eliminated---you now have to specify the first column as NaN yourself.
    E.g., Apop_col(data, -1, to_nan); gsl_vector_set_all(to_nan, GSL_NAN);
    This will make things more predictable, and save you if(!has_nans)... else... kind of statements.
**Removed the prepared element of the apop_model.
**apop_model_prep ==> apop_prep for consistency with other apop_model dispatch functions. apop_model_prep left for now as an alias in deprecated.h
!!apop_parameter_model: a method for getting the distribution of a parameter.
    **Moved OLS-family test stats (pval, qval, whatever) to a page of your_estimate->info. It won't be there for long either.
--settings macros let you use lowercase, thus entirely ignoring that they're macros.
    **apop_settings_rm_group function, which you were probably not using, changed to apop_settings_remove_group; apop_settings_get_group function => apop_settings_get_grp. Having a macro and a function that differed by a question of case was a bad idea to begin with.

    February 2010
!!Overhauling the output from estimations; pardon our dust. 
--Added CDF method to the apop_model, including apop_cdf dispatch method and default via random draws.
**Defininitvely removed the residual, covariance, and llikelihood elements from the
    apop_model struct. The first two will be pages appended to the data and parameters,
    respectively, and the last will be in the Info page appended to the parameters.
**Renamed apop_ls_settings (least square) to apop_lm_settings (linear model) "s/apop_ls/apop_lm/g" should work.
**Sundry lists of scalars, like the R^2 table and the estimation routine's info table put the data in column zero, not column -1. In the next bullet point you'll see how this simplifies retrieval.
**Added an info element to the apop_model--> more shuffling of auxiliary info. 
    --Find results like the log likelihood or AIC via, e.g., apop_data_get(your_model->info, .rowname="log likelihood");
  **Find the predicted/residuals via apop_data_get_page(your_model->info, "Predicted");
        This means that the input data set is read-only again. 
    --Find the parameter covariances via apop_data_get_page(your_model->parameters, "Covariance");
    **apop_estimate_coefficient_of_determination takes in the model again. Just replace est->parameters in your argument with est. apop_ols calls this fn automatically now [apop_data_get(your_model->info, .rowname="R sq")], so you probably aren't even calling it anymore.
**apop_data_add_named_elmt now writes to the zeroth element of the matrix, not the vector.
So instead of apop_data_get(data, .rowname="R squared", -1), just go with apop_data_get(data, .rowname="R squared"). This affects many of the elements of the info-type matrices.
--apop_data_pack, apop_data_unpack, apop_ml_impute, apop_map offer an option to use all pages. 

    0.23 January 2010
**expected_value element of the model renamed predict; made coherent across models.
!!apop_data set now has a ->more pointer to an additional apop_data set, e.g., for data + covariances or predictions + confidence intervals.
--apop_ml_imputation renamed to apop_ml_impute. "#define apop_ml_imputation apop_ml_impute" retains noun-form name, but consider it deprecated.
!!apop_estimate now copies, preps, then estimates. Estimate method of apop_model struct can thus assume the copy/prep step has been done;
probably should not do these itself. As a side-effect, apop_maximum_likelihood's second argument is now a apop_model* (used to be apop_model).
--apop_regex and apop_strcmp, for easier searching through your info pages.
--minor rewording of COPYING2.
**Because the Predicted table is now part of the parameter set, not the model,
    apop_estimate_coefficient_of_determination now takes in the parameter set, not the model. Just replace est in your argument with est->parameters.
**apop_multinomial_probit folded into apop_probit, where it should've been all along.
    Regex for the fix: "s/apop_multinomial_probit/apop_probit/g"

    December 2009
--apop_strcmp
--apop_loess model: 3,500 lines of code from the netlib archive, lovingly restored.

    November 2009
--apop_rm_columns bug fixed by Birger Baksaas.
--apop_text_to_db attaches numeric affinity to sqlite3 columns, making numeric comparisons easier.
--apop_histogram_model_reset's first argument is now "base" instead of "template", as a concession to C++ users.

	September 2009
--Many minor changes, mostly regarding adding optional arguments.
--Dirichlet model
--Output functions now take a consistent set of specs regarding to where they will write. You no longer have to use the global apop_opts settings if you don't want to.

	August 2009
--apop_map and apop_map_sum. Reworks the apop_(map|apply) system to be more flexible but a little more complex.
-apop_(data|matrix|vector)_fill is now more robust---no more int vs float issues.
**Removed apop_count_cols
--Default univariate RNG, if you don't have one: Adaptive rejection markov chain sampling
**.use_covar and other such settings now take 'y' or 'n', not 0 or 1.
--new macro Apop_settings_set = Apop_settings_add, but makes more human sense
--numeric covariance, formerly maligned, now works.

	July 2009
--multivariate gamma, log-gamma.
--t, F, chi^2, Wishart distributions, for description [and Bayesian
updating]
--apop_matrix_to_positive_semidefinite and apop_matrix_is_positive_semidefinite 
--bug fixes
--Apop_model_add_group replaces Apop_settings_add_group, and is much more easy to work with.

	June 2009
--More variadicized functions
	--notably, apop_estimate is much more useful
--apop_opts.version.
--apop_(vector|matrix|data)_stack have an inplace option, making stacking inside a for loop easier.
--apop_test convenience function
--more autoconf macros ==> some compilation hacks now done right

	May 2009
--mysql functions slightly cleaned up
--apop_opts.db_user and apop_opts.db_pass for mysql.
!!Functions that take lots of basically optional inputs, like apop_text_to_db, now use some designated initializer magic to let the user rearrange or omit inputs.
**apop_dot also now allows designated initializers, which breaks
(only) calls of the form apop_dot(a_vector, a_matrix, 't'). Replace with
apop_dot(a_vector, a_matrix, 0, 't') or apop_dot(a_vector, a_matrix,
.form2='t')
--With optional inputs, some functions now handle RNGs for the lazy user ==>added apop_opts.rng_seed
--apop_vector_distance is much more versatile
**Removed apop_matrix_summarize. Too much like apop_data_summarize. Just
replace every instance of apop_matrix_summarize(m) with apop_data_summarize(apop_matrix_to_data(m)).


    April 2009
--sample moments are now mega-accurate---possibly the most unbiased estimators in code today.
!!Python interface via swig

	March 2009
--apop_matrix_realloc, apop_vector_realloc
--sqlite queries no longer rely on a temp table ==> faster
--fixed bugs in apop_table_exists making queries fail in Cygwin.

	Jan/Feb 2009
--Added more tests; some cleanup in test.c
--Binomial distribution looks in both the data set's vector and matrix

	December 2008
--When writing x=infinity to a db table, I now write 'inf' to the db, instead of breaking.  SQLite has no standard here.

	October 2008
--bug fixes to new apop_data_show
--bug fixes to apop_bernoulli.p
--apop_update tweaks

	September 2008
--Documentation overhaul
--apop_data_show is much more screen-friendly. Keep using
apop_data_print for more machine-readable and less fixed-width output.
**apop_plot_histogram now takes in a vector, bin count, and name of output. This is what it did in the first half of the year.
  The current version of apop_plot_histogram, which acts on a histogram model, is renamed to apop_histogram_plot.

	August 2008
--Constraints in ML work better.
**Overhaul of some discrete choice models
	--Added tests for the probit and logit.
	--fixed a bug revealed by the tests
	**the first choice has a fixed value of zero.
	**You'll probably need to call Apop_category_settings_add before estimating 
		your model, unless the outcome choice variable is the 0th column of the matrix.
	--more to come, e.g., multinomial probit will be merged with ordered probit.
-Adding a settings group of a given type when that group already exists used to induce an error; now the old type is replaced with a clean default.
--bug fix for apop_test_fisher_exact on non-square matrices
--apop_settings_add and company do more work in functions and less in macros.
--removed the settings_ct element of the apop_model; using a sentinel at the end of the array instead.
--Slightly improved reading of text files.
--Bootstrap/jackknife act on models with parameters in both matrix and vector form.


    July 2008
--Guts of apop_plot_histogram now use the apop_histogram model. 
**  Also, it no longer normalizes the histogram to integrate to one by
  default. You need to explicitly request this via
  apop_histogram_normalize
--apop_plot finally deleted.
--apop_histogram_plot deleted; use apop_plot_histogram.
--Added apop_vector_skew_sample, apop_vector_kurtosis_sample.


	May 2008
--apop_settings_rm_group added.
--mysql interface has the beginnings of support for multiple
	semicolon-separated queries in one call.
--apop_histogram_refill_with_model ==> apop_histogram_model_reset;
	apop_histogram_refill_with_vector ==> apop_histogram_vector_reset.

	April 2008
--apop_dot handles names.
--apop_t_test now behaves correctly when one vector is of length 1.
--some improvements/fixes when dealing with mySQL.
--apop_sv_decomp renamed to apop_matrix_pca. Minor changes so that it correctly works as such.
--apop_text_to_(db|data) handles column names like it used to, which
   works better. Also a few other fixes for odd situations.

	March 2008
--Various improvements in reading in from text.
	-- a, "b, c", d will now correctly read in as three elements: a; then "b, c"; then d
	-- a,,b,c reads as a, NAN, b, c.

	February 2008
--Some of the header references didn't work for a fresh install.
--bug fixes, esp. with apop_test_kolmogorov
--added convenience fn apop_data_transpose

	January 2008
--Apop_assert, which streamlines the use of apop_error (thus shrinking the code base by 2%).
--apop_OLS now has a log likelihood (also shuffled some of the code around)
--bug fix in apop_binomial.p.
**More name reform: apop_correlation_matrix --> apop_matrix_correlation; apop_data_correlation_matrix --> apop_data_correlation; apop_covariance_matrix --> apop_matrix_covariance; apop_data_covariance_matrix --> apop_data_covariance.
--apop_count_(rows|cols)_in_text are now static functions and removed from the documentation.
--Removed apop_random_beta, which had been set as deprecated earlier.  Use apop_beta_from_mean_var.
**Removed apop_vector_isnan---just use apop_vector_map_sum(your_vector, isnan)
**Removed apop_vector_finite---just use apop_vector_bounded(your_vector, INFINITY)
--For your convenience, added Apop_settings_alloc() macro.
	**apop_histogram_params ==> apop_histogram_settings
	**apop_kernel_density_params ==> apop_kernel_density_settings
	[The settings/params distinction is in some ways arbitrary anyway.]
--bug fix in apop_mle.c: wasn't copying output parameters to the estimated model in some cases.
--As part of name reform, all function names are being switched to
lower-case throughout, so apop_ANOVA ==> apop_anova. Am keeping the old
forms via macros. Notice also the non-yelling macro capitalizations above, such as Apop_assert.
!!**Revised the settings for the apop_model. model_settings and
method_settings are out, replaced by a much more organized single list
of settings.
--added the apop_lookup command-line program.
**Renamed the apop_multinomial_logit model the apop_logit, because the
binary logit is a special case that requires no special handling.
**Reversed the signs on the probit coefficients, to better conform to
the norm.


	December 2007
--added apop_vector_moving_average
--apop_model now has prep and print methods.
**apop_p, apop_log_likelihood, apop_score now take a pointer to an apop_model, not the model itself.
--apop_multinomial_probit model
--When a data set has matrix and vector, apop_dot accepts a 'v' to use the vector.
--apop_plot_lattice produces a more attractive (and standard-form) plot.
--apop_data_print works much better now.
**apop_model no longer requires your data input to be const. It probably
will be const, but it's not the interface's place to dictate that.
**apop_data_unpack no longer allocates a new data set, but writes to an input data set assumed to be of the right size.
--added apop_ANOVA to produce one- or two-way ANOVA tables from the database.
**apop_test_ANOVA renamed to apop_test_ANOVA_independence to create a little more cognitive distance.
--apop_data_text_to_factors
--APOP_COL_T and APOP_ROW_T macros, to pull a column or row by name
--apop_beta_from_mean_var produces a beta-distributed model with the right (alpha, beta) parameters. 
**Thus, apop_random_beta is now marked as deprecated.
**apop_x_prime_sigma_x removed, on grounds of being silly. If you want it back, see model/apop_multivariate_normal.c, where it is now a static function.
**apop_qq_plot --> apop_plot_qq
--Multinomial logit model (and the probit now has names).
**Revised the bin-syncing methods. apop_vectors_to_histogram and apop_model_to_histogram are now out; apop_histogram_refill_with_vector and apop_histogram_refill_with_model are in. 
**Also removed apop_model_test_goodness_of_fit as redundant. Just produce a histogram, use the above refill functions, and send your two histograms to apop_histograms_test_goodness_of_fit. If you do this often, you can write a convenience function to do that as quickly as I could.
**apop_vector_replace and apop_matrix_replace are redundant---just use apop_(vector|matrix)_apply.
--The covariance matrix is now produced via the derivative of the score function at the parameter. I follow Efron and Hinkley in using the estimated information matrix---the value of the information matrix at the estimated value of the score---not the expected information matrix that is the integral over all possible data.

	November 2007
--added apop_model_copy_set_string to get a copy of a model whose
model_settings is just a string.
**Thanks to this, folded all of the _rank versions of models into their
base models. Set model_settings to "r" to use the rank version.
--The default MLE method is now the Nelder-Mead Simplex algorithm,
instead of the Fletcher-Reeves conjugate gradient. This is more
conservative.
--apop_(vector|matrix|matrix_all)_map_sum to get the sum of a function
applied to a vector. E.g., find count of NaNs with apop_vector_map_sum(v, isnan);
--apop_logit bug fix.

	October 2007

--apop_estimate now defaults to using MLEs, meaning you don't have to explicitly specify an estimate method for MLE models. 
--apop_crosstab_to_db reads both the matrix and text elements of the input apop_data set. 
--apop_system convenience function, to make C feel more like a scripting language. 
--Added some SQLite functions for mySQL compatibility: var_samp, var_pop, stddev_samp, stddev_pop, std. 
--Probit patched to not NaN for very unlikely parameter/data combinations. 
**apop_estimate_restart takes two models, rather than one model and some haphazard settings. 
--apop_plot_query no longer forces you to use the -d and -q switches to specify the database and query. 
**The two places that use regular expressions: apop_opts.db_nan and the search for a name via apop_data_get/set... use case-insensitive EREs. Before I'd been using BREs, which nobody likes. 
--<ctrl-c> stops the MLE searches, prints output, and continues the program. Especially useful for simulated annealing. [GDB tip: use the command: signal SIGINT ] 
--apop_mle_fix_params debugging: gradients work now. 
--apop_test_ANOVA added, to test the null hypothesis that all cells of a crosstab are equally likely. 
**apop_multivariate_normal_prob removed. Use the apop_multivariate_normal model and its .log_likelihood, .p, .draw, et cetera. 
**sed -i -e "s/apop_OLS_params/apop_ls_settings/g" -e "s/apop_mle_params/apop_mle_settings/g" *.c *.h

	September 2007
--The optimization methods now have an enumerated type.
**apop_opts.mle_trace_path is now the trace_path element of the apop_mle_params struct. Also, it works much better.
--apop_histogram_normalize function 
--improvements to apop_kernel_density and apop_histogram_print

	August 2007
--removed apop_model_template. Just copy one of the existing models. 
--apop_data_ptr_ti, apop_data_ptr_ii, apop_data_ptr_it 
--And you can never have too many bug fixes

	July 2007
--apop_binomial model takes two types of input now: a two-column form with hit count and miss count, and a list of binary hits or misses. 
--apop_lognormal model 
--bug fixes on Information matrix calculation.

	June 2007
[subversion ate this part; sorry.]

	May 2007
--apop_query_to_mixed_data
**apop_produce_dummies now makes dummy variables from both data and
text. This means that there's another parameter you need to set to 'd'
or 't' to indicate what you want dummified.
!!**Merged the apop_params and apop_model structures, leaving everything
in just one struct. That's about all the merging left.
--apop_text_alloc and apop_text_add to make text manipulation a little
easier.
--apop_matrix_apply_all and apop_matrix_map_all operate on all items in
a matrix.
**small tweak to apop_vector_normalize interface.
--apop_matrix_inverse and apop_matrix_determinant, because the
apop_det_and_inv interface is sort of ugly.
--APOP_SUBMATRIX macro
--MLEs put the expected score in the ->more element of the returned
apop_model. If people find this useful, we can maybe put a
proper expected_score element in the model.
--More consts in function headers. You can decide whether this
is actually useful.

	0.19 April 2007
!!**Eliminated the apop_estimate and apop_ep structures, replacing them
with the apop_params structure. The apop_params + apop_model pair form a
closure representing a parametrized model. Expect the uses element to go
away soon; after that, things should be stable. Parameters for individual
methods now have their own space; try apop_ml_params_alloc and
apop_OLS_params_alloc, for example.
If you are just doing things like
apop_estimate_show(apop_OLS.estimate(data,NULL));
then don't worry, but if you are doing a lot with the input parameters,
then have a look at Chapter 6 of _Modeling with Data_.
--apop_histogram model
**Gradually rewriting the histogram functions from before to make use of
that model. E.g., eliminated apop_vector_to_cmf.
**apop_line_to_data fixed to use both vector and matrix terms. Now
requires arguments: (indata, vsize, m1size, m2size).
--MLE now approximates the Information matrix using data gathered during
the MLE search. This is wrong but cheap; right but expensive procedures
forthcoming. [Hint: Simulated annealing gathers more info.]
--apop_(data|matrix|vector)_fill functions, which are a touch fragile,
but very useful when used with care.
**apop_data_(get|set)_(tn|nt) changed to (ti|it), because n could stand
for name or number, while i stands for index, and is often used for integers.
--apop_names now have a title element, so you can give your data
structures a title.
**apop_params_alloc takes an apop_model, not an &apop_model. It's more
natural that way.
**Finally erradicated every last vestige of inventories: apop_params no
longer has a .uses element. Instead, apop_specific_model_params may have
a want_cov, want_expected_value, want_whatever element if the element is
optional. And really, the parameters themselves should never be
optional. What was I thinking.
**apop_model_fix_params now sets up and returns an apop_mle_params object,
thus resolving the problem that the MLE params needs a model input,
and the model_fix_params model needed an MLE params input.

	0.18 March 2007
**apop_text_to_db now assumes column names unless you specify -nc.
--If you set parameter_ct==0 in your model definition, the MLEs will
assign parameter_ct == the number of columns in your data set.
--Missing data functions: apop_listwise_delete and apop_ml_imputation
**The constraint element of the apop_model now takes a void* parameter,
like it should have all this time.
**apop_jackknife (1) renamed to apop_jackknife_cov and (2) now actually
works.
**Entirely eliminated the apop_inventory structure. Its sole utility is
inside the apop_ep struct.
**Changed the RNG interface for the sake of allowing multidimensional
draws. [Not that I have any functions that do that right now.]
--Bayes-oriented MCMC algorithm: apop_mcmc_update
!!Bayesian model generator: apop_update
**apop_model paramters is now an apop_model. See the documentation of
the model for all the changes.
**apop_data_alloc now takes three arguments: vsize, msize1, msize2. To
update, just put a 0, at the head of the arg list.
!!An absolutely fabulous apop_linear_constraint function.
--Produce a model with some parameters fixed via apop_model_fix_params.
--apop_beta model

	February 2007

**Apop_sv_decomposition has a slightly nicer interface.
**data->categories was too much to type, and too specific. The apop_data
struct now has a data->text element, and textsize[0] and [1]. The
categories element is linked to this, but is now deprecated.

	January 2007
**Root finding hooked into the max likelihood fns.

0.17	December 2006
**Apop_model struct has lost the fdf object, which was annoying, and now
has the p function.
--mySQL support.
**apop_query_to_chars now returns an apop_data structure, so you don't
have to go back and gather column names and dimensions.
**apop_name_get (and the apop_get_tt family) now use regular expressions
instead of SQL's LIKE operator. This is _much_ faster.
--the apop_distribution model.

		November 2006
--The preeminently useful APOP_COL and APOP_ROW
--apop_data_calloc
--apop_vector_(apply|map) debugged.
**apop_estimation_params is just too darn long; reduced to apop_ep.


		October 2006
--apop_text_to_db now reads from STDIN.
**deleted apop_query_db; use apop_query.
--Kolmogorov-Smirnov test.
--apop_t_test returns GSL_NAN when given a one-element vector instead of
hanging.
--no more soft links in the tgz file==>may work better on Windows machines.
--apop_(vector|matrix)_(map|apply) will apply a function to every
row of a matrix or every element of a vector. The map functions return a
gsl_vector.
--bug fix in apop_test_goodness_of_fit.

		September 2006

**Removed apop command-line server thing. It was interesting, but that's
the best that could be said of it.
--Added functions for weighted data: weighted least squares, weighted moments.
--apop_vector_percentiles now allows for averaging instead of rounding.

		August 2006
**apop_log_likelihood and friends now demand that data be apop_data*,
rather than void*. Too many things broke when users gave non-apop_data
data.
--bug fixes

		July 2006
--Many more checks for NULL ==> more robust code and easier debugging.
--Bug fixes.
**apop_data_split, and apop_data_stack has been revised to handle the
idea that a vector is the -1st element of the matrix. I.e., check your
code if you're trying to merge matrices without merging the vectors.
--Lattice plots.
--Convenience t-tests from inside model estimations are fixed.
--apop_query_to_vector. 
--apop_opts.output_type == 'p' to print to apop_opts.output_pipe
**apop_..._print and apop_..._show now work out whether elements are
integers (if (val == (int) val)...), and print accordingly. This means
that apop_..._print_int and apop_..._show_int are basically obsolete,
and have been removed.
--Apop_OLS now allows weights.
--Test library now includes a few NIST certified tests.

		June 2006
--added preprocessor cruft to let the library work for C++
--Jackknife revised

		May 2006
**The apop_model no longer includes an inventory. I leave it to the
estimate function to do its own allocation.

		April 2006
**apop_matrix_normalize and apop_vector_normalize had different
numbers for the same normalizations. Was that ever dumb. Also, I've
switched to chars instead of ints to signify this stuff, for better
mnemonics without resorting to the
APOP_ENUM_YOU_HAVE_TO_LOOK_UP_EVERY_TIME_BECAUSE_ITS_SO_LONG sort of
thing. If you were using apop_matrix_normalize(data, 0) before, you
need to change that to using apop_matrix_normalize(data, 'm'). Thus,
apop_vector_normalize now has one more normalization, for a total of
four for both.

--Added apop_rng_alloc convenience fn.

--Added apop_strip_dots to keep inputs to the database healthy.

--apop_name_find uses LIKE instead of strcmp.

--a fn to calculate the generalized harmonic.

--A whole section on histograms and goodness-of-fit tests.

--apop_data_set fns to go with the apop_data_get fns.

--apop_data now includes a vector type
	--apop_estimate.parameters is now an apop_data type.
	--apop_estimate.names is thus obsolete.

		March 2006
**apop_inventory is now a subset of apop_estimation_params. Implications:
	--added apop_estimation_params_alloc() to ensure that inventory is set right.
	--the model.estimate(data, inv, params) method is now model.estimate(data, params)
	  model.estimate(data, NULL) still does what the user expects it to.
  This makes structural sense, but will lightly break any existing code.
  fix: change 
   apop_inventory *inv = apop_inventory_set(1);
   model.estimate(data, inv, NULL);

   to
   apop_estimation_params *ep = apop_estimation_params_alloc();
   model.estimate(data, ep);

   and in any apop_estimates, change any use of est->uses to
   est.estimation_params.uses.

**Next apop_estimate reform: y_values and residuals combined into one
apop_data table with actual, predicted, residual columns.
	--obviated the need for a 'dependent' element in apop_names; removed that.
	If you need the name, it's now your_est->dependent->names->colnames[0].

**your_estimate->covariance is now an apop_data set instead of a gsl_matrix.

**the data element of the apop_matrix structure is now named matrix. So
instead of data_set->data, use data_set->matrix, and instead of
estimate->data->data->data, you can use estimate->data->matrix->data.

--The command-line utility has been revisited, and can do a few more
things, like OLS.

--Simulated annealing
	--added convenience fns apop_vector_distance(pt1,pt2) and
		apop_vector_grid_distance(pt1,pt2)

**Apop_data_memcpy no longer malloc()s for you, for comparability with
the world's other memcpy fns. If you want mallocing, use apop_data_copy.

--apop_test_fisher_exact(). Cut 'n' pasted from R, who cut 'n' pasted it
from somebody else. Despite being the same code, it runs fifty (50)
times faster from Apophenia.

		February 2006
--sort-of-adaptive MLE: use apop_estimate_restart to execute a new MLE
search beginning where the last one ended, perhaps using a new method or
rescaled parameters.
	--This needed convenience functions to check for divergence, thus
		added apop_vector_finite, apop_vector_bounded, apop_vector_isnan.
**apop_db_to_crosstab now returns an apop_data set instead of a gsl_matrix.
Also, it finally works with column headers that aren't numeric.
**stats like apop_mean are now apop_vector_mean, following the proper
	pkg-noun-verb naming scheme. 
--Textbook is much improved.
--apop_vector_to_pdf convenience fn.

--Some of the fns that used to be of the form
	apop_get_something(input, &output);
are now of the more natural
	out	= apop_get_something(input);
This includes apop_array_to_vector and apop_array_to_matrix

--bootstrapping works, and works with with apop_models.
--apop_poisson model

0.15	January 2006
Added an apop_opts structure for options. Alowed the following changes:
	--apop_verbose  is now apop_opts.verbose. Try this on your existing code:
		perl -pi.bak -e 's/apop_verbose/apop_opts.verbose/g' *.c *.h
	--the output functions now output into three formats: on screen, to file, to db;
		see chapter five of the manual.
		
--F tests 
--R squared.

0.14	December 2005

The apop_data structure, which is just a shell for a gsl_matrix and an
apop_name. Was just sick of sending names following around my tables. 
Lets us keep both numerical and categorical data in one place; kind of
like R's data frame.

--Added linear model objects: OLS, GLS. This means that what had been
the apop_OLS function is now the apop_estimate_OLS function, and where it
used to take in a gsl_matrix and an apop_name, now it takes an apop_data
structure and a NULL. So you'll have to modify your code accordingly.

--A function to generate dummy variables, useful in conjunction with
--Functions to stack matrices and apop_data sets. Even a
apop_partitioned_OLS function, that will only practically work for small data sets.

--pow(.,.) in the database. I can't believe I dealt with SQL this long w/o it.

0.13	December 2005

--The apop_model object. This was a big deal that deserves more than
just one line; see the manuals.

0.12 	mid November 2005

--Bar charts (assuming you've got Gnuplot > 4.1)
--percentiles (in case you haven't got it)
**redid MLE system so you can pick among the many options now available.  As a part of this:
	--better handling of constraints.
	--numerical gradients.
	--numerical Hessians.

0.12 	early November 2005, post-hiatus

--You now have three maximum likelihood estimators to choose from: the
GSL's no-gradient, the GSL's with-gradient, and Mr. WN's autocalculated
gradient. 

--If you haven't seen it before, the apop_distribution structure is
increasingly well-supported. It allows the user to specify the features
of the Max. Likeihood model in a consistent manner which facilitates
things like comparing two models.

--I'd still suggest taking the Waring and Yule distributions with care;
everything else seems to check out.


0.12 	September 2005
**The distributions are now objects, which just provides a neat way
of grouping together the half-dozen functions which are associated with
any one distribution.

0.11	September 2005
--command-line server is much improved. I actually do work with it.
--Documentation is now via doxygen.
--asst bug fixes.
--Have started to take plotting (via gnuplot) seriously
--a limited test suite. Try: make test .

0.10	August 2005
--This version includes a server to park itself in memory and receive data
processing requests. The intent is that one can then do analysis from
the command line or a Perl/Python/Whatever script. The client/server
works in the sense of handling a handful of requests without
segfaulting, but remains in proof-of-concept stage. 
--Added apop_merge_db for joining databases, both via C and command line 
--Run t tests from the cmd line or the database.

0.09	July 2005
--Flattened the relatively complex vasprintf subsystem from GNU, so if
you've been having trouble compiling on non-GNU systems, try again.
Added two little command-line programs. Also, added more little
functions which aren't very interesting, like t-tests; maybe you'll
stumble upon them.

0.08	May 2005
--OLS/GLS/MLE now properly support the apop_estimate structure 
--Column names

0.07	April 2005
--uses the apop_estimate structure to return heaps of data from regressions & MLEs
--uses the apop_model structure

0.06
--var(x), skew(x), kurtosis(x) added to SQL understood by Apophenia.

0.05
--added a little crosstab utility
--queries now accept printf-type arguments. 
	==>GNU vasprintf was added.
		==>updated to work with autoconf 1.7